REVERSIBLE INTEGER-TO-INTEGER WAVELET TRANSFORMS FOR IMAGE CODING by MICHAEL DAVID ADAMS B.A.Sc, The University of Waterloo, 1993 M.A.Sc, The University of Victoria, 1998 A THESIS SUBMITTED IN PARTIAL FULFILMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY in THE FACULTY OF GRADUATE STUDIES (Department of Electrical and Computer Engineering) We accept this thesis as conforming to the required standard THE UNIVERSITY OF BRITISH COLUMBIA September 2002 © Michael David Adams, 2002 In presenting this thesis in partial fulfilment of the requirements for an advanced degree at the University of British Columbia, I agree that the Library shall make it freely available for reference and study. I further agree that permission for extensive copying of this thesis for scholarly purposes may be granted by the head of my department or by his or her representatives. It is understood that copying or publication of this thesis for financial gain shall not be allowed without my written permission. Department of Electrical and Computer Engineering The University of British Columbia Vancouver, Canada Date Sepi. ? . Q 7 2oO? 11 Ab s t r a c t Reversible integer-to-integer (ITI) wavelet transforms are studied in the context of image coding. Considered are matters such as transform frameworks, transform design techniques, the utility of transforms for image coding, and numerous practical issues related to transforms. The generalized reversible ITI transform (GRITIT) framework, a single unified framework for reversible ITI wavelet/block transforms, is proposed. This new framework is then used to study several previously proposed frame-works and their interrelationships. For example, the framework based on the overlapping rounding transform is shown to be a special case of the lifting framework with only trivial extensions. The applicability of the GRITIT framework for block transforms is also demonstrated. Throughout all of this work, particularly close attention is paid to rounding operators and their characteristics. Strategies for handling the transformation of arbitrary-length signals in a nonexpansive manner are considered (e.g., symmetric extension, per-displace-step extension). Two families of symmetry-preserving transforms (which are compatible with symmetric extension) are introduced and studied. We characterize transforms belonging to these families. Some new reversible ITI structures that are useful for constructing symmetry-preserving transforms are also proposed. A simple search-based design technique is explored as means for finding effective low-complexity transforms in the above-mentioned families. In the context of image coding, a number of reversible ITI wavelet transforms are compared on the basis of their lossy compression performance, lossless compression performance, and computational complexity. Of the transforms considered, several were found to perform particularly well, with the best choice for a given application depending on the relative importance of the preceding criteria. Reversible ITI versions of numerous transforms are also compared to their conventional (i.e., non-reversible real-to-real) counterparts for lossy compression. At low bit rates, reversible ITI and conventional versions of transforms were found to often yield results of comparable quality. Factors affecting the compression performance of reversible ITI wavelet transforms are also presented, supported by both experimental data and theoretical arguments. In addition to this work, the JPEG-2000 image compression standard is discussed. In particular, the JPEG-2000 Part-1 codec is described, analyzed, and evaluated. iii Conten ts Abstract ii Contents iii List of Tables vii List of Figures viii List of Algorithms x List of Acronyms xi Preface xiii Acknowledgments xiii Dedication xv 1 Introduction 1 1.1 Reversible Integer-to-integer (ITI) Wavelet Transforms 1 1.2 Utility of Reversible ITI Wavelet Transforms 1 1.3 Common Misconceptions 2 1.4 Historical Perspective 2 1.5 Overview and Contribution of the Thesis 4 2 Preliminaries 6 2.1 Introduction 6 2.2 Notation and Terminology 6 2.3 Multirate Filter Banks and Wavelet Systems ' 8 2.3.1 Multirate Systems 9 2.3.2 Sampling 9 2.3.3 Downsampling 9 2.3.4 Upsampling 11 2.3.5 Noble Identities 11 2.3.6 Polyphase Representation of Signals and Filters 11 2.3.7 Filter Banks 13 2.3.8 Uniformly Maximally-Decimated (UMD) Filter Banks 13 2.3.9 Perfect-Reconstruction (PR) UMD Filter Banks 13 2.3.10 Polyphase Form of a UMD Filter Bank 14 2.3.11 Conditions for PR System 16 2.3.12 Octave-Band Filter Banks 16 2.3.13 UMD Filter Bank Implementation 17 2.4 Image Codi ng 17 2.4.1 Image Coding and Compression 17 2.4.2 Transform-Based Image Compression Systems 18 2.4.3 Wavelet Transforms for Image Compression 19 2.4.4 Compression Performance Measures 19 Contents iv 3 Frameworks for Reversible ITI Wavelet/Block Transforms 21 3.1 Introduction 21 3.2 Rounding Operators 22 3.2.1 Integer-Bias Invariance and Oddness 22 3.2.2 Relationships Involving Rounding Functions 23 3.2.3 Rounding Error . 25 3.3 Previously Proposed Frameworks 28 3.3.1 S Transform 28 3.3.2 S+P Transform Framework 29 3.3.3 Lifting Framework 30 3.3.4 Overlapping Rounding Transform (ORT) Framework 32 3.4 Generalized Reversible ITI Transform (GRITIT) Framework 33 3.4.1 Primitive Reversible ITI Operations 35 3.4.1.1 Split Operation 35 3.4.1.2 Join Operation 35 3.4.1.3 Displace Operation 35 3.4.1.4 Exchange Operation 36 3.4.1.5 Shift Operation 37 3.4.1.6 Scale Operation 37 3.4.2 Reversible ITI Wavelet Transforms 39 3.4.3 Reversible ITI Block Transforms 39 3.4.4 Practical Considerations Concerning Rounding 40 3.4.5 GRITIT Realization of Wavelet/Block Transforms 42 3.4.6 Variations on the GRITIT Framework 43 3.5 Relationship Between GRITIT and Other Frameworks 44 3.5.1 S Transform 44 3.5.2 S+P Transform Framework 45 3.5.3 Lifting Framework 45 3.5.4 ORT Framework 45 3.6 Relationship Between the ORT and Lifting Frameworks 48 3.7 Generalized S Transform 50 3.7.1 Block S Transform 51 3.7.2 Generalized S Transform . . . 51 3.7.3 Calculation of GST Parameters 52 3.7.4 Choice of Rounding Operator 55 3.7.5 Examples 57 3.7.6 Practical Application: Modified Reversible Color Transform 58 3.8 Summary 59 4 Nonexpansive Reversible ITI Wavelet Transforms 61 4.1 Introduction 61 4.2 Extension Methods 62 4.2.1 Periodic Extension 62 4.2.2 Symmetric Extension 62 4.2.3 Per-Displace-Step Extension 64 4.3 Symmetry-Preserving Transforms 64 4.3.1 Transform Families 65 4.3.1.1 OLASF Family 66 4.3.1.2 ELASF Family 66 4.3.2 Transforms and Symmetric Extension 67 4.3.2.1 OLASF Case 67 4.3.2.2 ELASF Case 68 4.3.3 Symmetry Preservation in the ELASF Base Filter Bank 69 4.3.3.1 Symmetry Preservation 70 Contents v 4.3.3.2 Modifying the E L A S F Base Filter Bank 70 4.3.4 O L A S F Family '. 73 4.3.5 E L A S F Family 74 4.3.5.1 Transform Properties 74 4.3.5.2 Lowpass Analysis Filter Transfer Function 77 4.3.5.3 Highpass Analysis Filter Transfer Function 77 4.3.5.4 Analysis Filter Lengths 78 4.3.5.5 Incompleteness of Parameterization 78 4.3.5.6 Analysis Filter Gains 78 4.3.5.7 Swapping Analysis and Synthesis Filters 79 4.3.6 Relationship Between Symmetric Extension and Per-Displace-Step Extension 79 4.4 Design of Low-Complexity Symmetry-Preserving Transforms 80 4.4.1 Transforms 80 4.4.2 Design Method 81 4.4.3 Design Examples 81 4.4.4 Coding Results 81 4.5 Summary 82 5 Reversible ITI Wavelet Transforms for Image Coding 86 5.1 Introduction 86 5.2 Transforms 87 5.3 Computational Complexity and Memory Requirements 88 5.4 Experimental Results 93 5.4.1 Evaluation Methodology 93 5.4.2 Lossless Compression Performance 95 5.4.3 PSNR Lossy Compression Performance 97 5.4.4 Subjective Lossy Compression Performance 97 5.4.5 Reversible ITI Versus Conventional Transforms for Lossy Compression 97 5.5 Analysis of Experimental Results 101 5.5.1 Reversible ITI Versus Conventional Transforms for Lossy Compression 101 5.5.1.1 Impact of 1IR Filters 101 5.5.1.2 Number of Lifting Steps 101 5.5.1.3 Rounding Function 103 5.5.1.4 Depth of Image 103 5.5.1.5 Bit Rate 103 5.5.2 Factors Affecting Compression Performance 103 5.5.2.1 Parent Linear Transform 103 5.5.2.2 Approximation Behavior 104 5.5.2.3 Dynamic Range 104 5.6 Summary 107 6 Conclusions and Future Research 108 6.1 Conclusions 108 6.2 Future Research 109 6.3 Closing Remarks 109 Bibliography 110 A J P E G 2000: A n International Standard for Still Image Compression 118 A . l Introduction 118 A.2 JPEG 2000 119 A.2.1 Why JPEG 2000? 119 A.2.2 Structure of the Standard 119 A.3 JPEG-2000 Codec 120 A.3.1 Source Image Model 120 Contents vi A.3.2 Reference Grid . 120 A.3.3 Tiling 121 A.3.4 Codec Structure 123 A.3.5 Preprocessing/Postprocessing 123 A.3.6 Intercomponent Transform 123 A.3.7 Intracomponent Transform 124 A.3.8 Quantization/Dequantization 127 A.3.9 Tier-1 Coding 1 2 8 A.3.10 Bit-Plane Coding 129 A.3.10.1 Significance Pass 130 A.3.10.2 Refinement Pass . 130 A.3.10.3 Cleanup Pass 130 A.3.11 Tier-2 Coding 132 A.3.11.I Packet Header Coding 133 A.3.11.2 Packet Body Coding 134 A.3.12 Rate Control 134 A.3.13 Region of Interest Coding 135 A.3.14 Code Stream 136 A.3.15 File Format 137 A.3.16 Extensions 138 A.4 Codec Evaluation 139 A.4.1 Evaluation Methodology 139 A.4.2 Code Execution Profiling 139 A.4.2.1 Lossless Coding 140 A.4.2.2 Lossy Coding 140 A.4.3 JPEG 2000 vs. Other Methods 140 A.4.3.1 JPEG 2000 vs. JPEG LS 142 A.4.3.2 JPEG 2000 vs. JPEG 143 A.4.3.3 JPEG 2000 vs. SPIHT 143 A.5 Summary 144 A.6 JasPer 145 Index 146 vii List o f Tables 3.1 Error characteristics for various rounding operators 25 3.2 Sets of predictor coefficients for the S+P transform framework 30 3.3 Luminance error for the M R C T and RCT 59 4.1 Transforms 82 4.2 Transform parameters 82 4.3 Lossy compression results 83 4.4 Lossless compression results 83 5.1 Transforms 87 5.2 Forward transforms 89 5.3 Transform parameters 90 5.4 Computational complexity of transforms 90 5.5 Test images 94 5.6 Lossless compression results 96 5.7 Relative lossless compression results 96 5.8 PSNR lossy compression results 98 5.9 Subjective lossy compression results (first stage) 98 5.10 Subjective lossy compression results (second stage) 98 5.11 Difference in PSNR performance between reversible ITI and conventional transforms 100 5.12 Difference in subjective performance between reversible ITI and conventional transforms 101 5.13 Influence of the number of lifting steps on lossless compression performance 104 5.14 Transform parameters affecting compression performance 106 5.15 Average coefficient magnitudes and lossless compression performance 106 A. 1 Parts of the standard 119 A.2 Types of marker segments 137 A.3 Box types 138 A.4 Test images 139 A.5 Codec execution profile for lossless coding 141 A.6 Codec execution profile for lossy coding 141 A.7 Comparison of JPEG 2000 and other methods for lossless coding 142 A.8 Comparison of JPEG 2000 and other methods for lossy coding 143 viii List of Figures 2.1 Examples of 2D lattices 10 2.2 Downsampler 10 2.3 Upsampler 11 2.4 Noble identities 12 2.5 Polyphase form of a filter 12 2.6 Analysis bank 13 2.7 Synthesis bank 13 2.8 Canonical form of a UMD filter bank 14 2.9 Polyphase realization of a UMD filter bank before simplification 15 2.10 Polyphase realization of a UMD filter bank 16 2.11 UMD filter bank (revisited) 17 2.12 Block cascade realization of the analysis polyphase matrix 18 2.13 Block cascade realization of the synthesis polyphase matrix 18 2.14 General structure of a transform-based image compression system 18 3.1 Identities for an integer-bias invariant rounding operator 22 3.2 S transform 28 3.3 S+P family of transforms 29 3.4 Lifting realization of a linear ID two-band wavelet transform 31 3.5 Two ladder steps 31 3.6 Modifying a lifting step to map integers to integers 32 3.7 Lifting realization of a reversible ITI ID two-band wavelet transform 32 3.8 Polyphase filtering networks for the lifting framework 33 3.9 ORT realization of a reversible ITI two-band wavelet transform 33 3.10 Polyphase filtering networks for the ORT framework 34 3.11 Operations for the GRITIT framework 38 3.12 General structure of a reversible ITI wavelet transform in the GRITIT framework 39 3.13 General structure of a reversible ITI block transform in me GRITIT framework 40 3.14 Combining multiple displace operations 42 3.15 Equivalent GRITIT networks for the networks of the ORT framework 46 3.16 Identities for interchanging D-and L-type operations 49 3.17 Identities for interchanging L-and S-type operations 50 3.18 BlockS transform 51 3.19 Realization of the generalized S transform 52 3.20 Realization of the C and C~x matrices 55 3.21 Equivalent representation of the generalized S transform for an integer-bias invariant rounding operator 56 4.1 Periodic extension of a signal 63 4.2 Types of signal symmetries 63 4.3 Symmetric extension examples 64 4.4 Displace step in the two-channel case 65 4.5 Displace step with PDS extension 65 4.6 Lifting realization of a reversible ITI wavelet transform 66 4.7 Two different reversible ITI approximations of the 6/14 transform 67 4.8 Base analysis filter bank for the ELASF family 70 List of Figures ix 4.9 General structure of a two-channel UMD filter bank 70 4.10 Linear version of the base analysis filter bank for the ELASF family 71 4.11 Modified base analysis filter bank for the ELASF family 71 4.12 Network consisting of a ladder step and rounding unit 71 4.13 Network consisting of an adder and rounding unit 71 4.14 Linear version of the base analysis filter bank for the new ELASF-like family 73 4.15 Modified base analysis filter bank for the new ELASF-like family 73 4.16 Lossy compression example 84 5.1 Synthesizing scaling and wavelet functions for the 5/3 transform 91 5.2 Synthesizing scaling and wavelet functions for the 2/6 transform 91 5.3 Synthesizing scaling and wavelet functions for the 9/7-M transform 91 5.4 Synthesizing scaling and wavelet functions for the 2/10 transform 91 5.5 Synthesizing scaling and wavelet functions for the 5/11-C transform 91 5.6 Synthesizing scaling and wavelet functions for the 5/11-A transform 91 5.7 Synthesizing scaling and wavelet functions for the 6/14 transform 92 5.8 Synthesizing scaling and wavelet functions for the 13/7-T transform 92 5.9 Synthesizing scaling and wavelet functions for the 13/7-C transform 92 5.10 Synthesizing scaling and wavelet functions for the 9/7-F transform 92 5.11 Synthesizing scaling and wavelet functions for the SPB transform 92 5.12 Synthesizing scaling and wavelet functions for the SPC transform 92 5.13 Lossy compression example 99 5.14 Lossy compression example 102 5.15 Lossy compression example 102 5.16 Subband structure for the 2D wavelet transform 105 A.1 Source image model 120 A.2 Reference grid 121 A.3 Tiling on the reference grid 122 A.4 Tile-component coordinate system 122 A.5 Codec structure 123 A.6 Lifting realization of a ID two-channel UMD filter bank 125 A.7 Subband structure 126 A.S Partitioning of a subband into code blocks 128 A.9 Templates for context selection 129 A.10 Sample scan order within a code block 131 A. 11 Partitioning of a resolution into precincts 133 A. 12 Code block scan order within a precinct 134 A. 13 Marker segment structure 136 A. 14 Code stream structure 136 A. 15 Box structure 137 A. 16 File format structure 138 A.17 Lossy compression example comparing JPEG 2000 and JPEG 144 X List of Algorithms A.I Significance pass algorithm 130 A.2 Refinement pass algorithm 130 A.3 Cleanup pass algorithm 131 A.4 Packet header coding algorithm 134 A.5 Packet body coding algorithm 135 List of Acronyms ID one dimensional 2D two dimensional bpp bits per pixel ASR arithmetic shift right BR bit rate CR compression ratio DC direct current D C T discrete cosine transform DFT discrete Fourier transform EBCOT embedded block coding with optimal truncation ELASF even-length analysis/synthesis filter EZW embedded zerotree wavelet FIR finite-length impulse response FPT forward polyphase transform GST generalized S transform ICT irreversible color transform IIR infinite-length impulse response IPT inverse polyphase transform ITI integer to integer ITU-T International Telecommunication Union Standardization Sector ISO International Organization for Standardization JBIG Joint Bi-Level Image Experts Group JPEG Joint Photographic Experts Group MAE mean absolute error MIMO multi-input multi-output MRCT modified reversible color transform MSE mean squared error NBR normalized bit rate List of Acronyms xii ORT overlapping rounding transform OLASF odd-length analysis/synthesis filter PAE peak absolute error PDS per displace step PR perfect reconstruction PSNR peak-signal-to-noise ratio RAFZ round away from zero RCT reversible color transform RGB red green blue ROI region of interest SISO single-input single-output SPIHT set partitioning in hierarchical trees TCQ trellis coded quantization UMD uniformly maximally-decimated VM verification model xiii Pre face It was the best of times, it was the worst of times, it was the age of wisdom, it was the age of foolishness, it was the epoch of belief, it was the epoch of incredulity, it was the season of Light, it was the season of Darkness, it was the spring of hope, it was the winter of despair... —Charles Dickens, A Tale of Two Cities, 1859 It is with mixed feelings that I reflect upon my years of study at the University of British Columbia (UBC). During this time, I was occasionally blessed with some very good fortune, and for this I am grateful. For the most part, however, these years proved to be, without a doubt, the most difficult of my life, both academically and personally. In these years, I experienced some of the highest highs and lowest lows. It was a time of the superlative degree. Hence, the above epigraph, the opening passage from Dickens' "A Tale of Two Cities", somehow seems apropos. Another interesting parallel of arguably greater literary relevance can be drawn between the above passage and my academic life at UBC. Those familiar with the tribulations of my studies may see this other parallel. Acknowledgments Certainly, help and encouragement from others are always appreciated, but in difficult times, such magnanimity is valued even more. This said, this thesis would never have been written without the generous help and support that I received from numerous people along the way. I would now like to take this opportunity to express my sincerest thanks to these individuals. First and foremost, I would like to thank my Ph.D. advisor, Dr. Rabab Ward. Rabab has always been kind, patient, and supportive. She has provided me with an environment conducive to learning and quality research. Perhaps, most of all, I would like to thank Rabab for always having confidence in me and my abilities, even when I did not. Next, I would like to express my gratitude to Dr. Faouzi Kossentini. For the first part of my Ph.D. studies, I was under Faouzi's supervision. In particular, I would like to thank Faouzi for having afforded me the opportunity to participate in JPEG-2000 standardization activities. I am also grateful to the following individuals, all of whom helped me in some significant capacity along the way: Angus Livingstone, David Jones, Michael McCrodan, Guy Dumont, Dave Tompkins, Michael Davies, Hongjian Shi, Mariko Morimoto, Sachie Morii, Emiko Morii, Yuen-Mui Chou, Sheri Chen, Sung-Suk Yu, Alice Yang, and G.C.. I would like to thank Evelyn Lau for writing "Runaway: Diary of a Street Kid". In her book, Evelyn documents the two harrowing years of her teenage life spent living on the streets of Vancouver and her struggle with drug addiction and prostitution. Her story moved me unlike anything that I have ever read before in my life. Her story has also been a great source of inspiration to me. Against seemingly insurmountable odds, Evelyn was able to, at least partially, triumph over her problems. Her words have helped me to find strength, courage, and faith in myself at times when such did not come easily. I would like to acknowledge the financial support of the Canadian federal government through a Natural Sciences and Engineering Research Council Postgraduate Scholarship. This support certainly helped to greatly reduce the financial burden associated with my studies. Lastly, I would like to thank two individuals without whom I might never have started my doctorate. First, I would like to thank my Master's advisor, Dr. Andreas Antoniou. It was Andreas that initially sparked my interest in digital signal processing, leading me to pursue research in this area. He also first introduced me to the world of research publications, and I know that my technical writing skills have benefitted substantially as a result of his many Preface xiv comments. Second, I would like to express my gratitude to Dr. Wu-Sheng Lu for having taught me my first course on wavelets and filter banks. I have never met a professor who is more enthusiastic about teaching. He has been a constant source of inspiration to me, and I hope that, one day, I can be half as good a teacher as he. Michael Adams Vancouver May 2002 To those who inspired me and gave me hope, even in my darkest hour Chapter 1 Introduction i Maelstrom Within Here I sit, my eyes transfixed upon the barren white canvas before me, emotions struggling to escape the bondage of my troubled mind, begging for release as a river of black ink upon the page. Yet, the words are elusive, taunting and teasing me, as they run dizzying circles in my mind. I clutch desperately for them, but their nimble bodies evade my touch, continuing their seductive dance just out of reach. Am I forever doomed to this life of torment, never to be free from the emotions raging within? Unable to bleed my feelings upon the page, am I destined to be consumed by them? —Michael D. Adams 1.1 Reversible Integer-to-integer (ITI) Wavelet Transforms In recent years, there has been a growing interest in reversible integer-to-integer (ITI) wavelet transforms for image coding applications [1,9,13,23,26,29,39,40,51,50,59,79,85,S6,87, 100,101,1 14,118,120,126,127,148,11,72,27, 102,44]. Such transforms are invertible in finite-precision arithmetic (i.e., reversible), map integers to integers, and approximate linear wavelet transforms. Due largely to these properties, reversible ITI wavelet transforms are extremely useful for image compression systems requiring efficient handling of lossless coding, minimal memory usage, or low computational complexity. For example, these transforms are particularly attractive for supporting functionalities such as progressive lossy-to-lossless recovery of images [ 11.4,1,112,14 8], lossy compression with the lossless reproduction of a region of interest [102,26], and strictly lossy compression with minimal memory usage [9], to name but a few. 1.2 Utility of Reversible ITI Wavelet Transforms One might wonder why reversible ITI wavelet transforms are such a useful tool. As suggested above, both the re-versible and ITI properties are of great importance. In what follows, we will explain why these properties are desirable. Also, we will briefly note a few other characteristics of reversible ITI wavelet transforms that make them well suited for signal coding applications. In order to efficiently handle lossless coding in a transform-based coding system, we require transforms that are invertible. If the transform employed is not invertible, the transformation process will typically result in some information loss. In order to allow lossless reconstruction of the original signal, this lost information must also be coded along with the transform data. Determining this additional information to code, however, is usually very costly in terms of computation and memory requirements. Moreover, coding this additional information can adversely affect compression efficiency. Thus, invertible transforms are desired. Often the invertibility of a transform depends on the fact that the transform is calculated using exact arithmetic. In practice, however, finite-precision arithmetic is usually employed, and such arithmetic is inherently inexact due to rounding error. Consequently, we need transforms that are reversible (i.e., invertible in finite-precision arithmetic). Fortunately, as will be demonstrated in later chapters, it is possible to construct transforms that are not only invertible, but reversible as well. In order to be effective for lossless coding, it is not sufficient that a transform simply be reversible. It must also yield transform coefficients that can be efficiently coded. In the lossless case, all of the transform coefficient bits must be coded to ensure perfect reconstruction of the original signal. Thus, we would like to avoid using any transform that Chapter 1. Introduction 2 unnecessarily introduces more bits to code. Assuming that the image is represented by integer-valued samples (which is the case in the vast majority of applications), we would like a transform that does not introduce additional bits to the right of the radix point. In other words, we would like a transform that maps integers to integers. If the transform were to introduce bits to the right of the radix point, many more bits would need to be coded in the lossless case, leading to very poor compression efficiency. The use of ITI transforms has other benefits as well. Typically, in the case of ITI transforms, a smaller word size can be used for storing transform coefficients, leading to reduced memory requirements. Also, this smaller word size may potentially reduce computational complexity (depending on the method of implementation). Reversible ITI wavelet transforms approximate the behavior of their parent linear transforms, and in so doing inherit many of the desirable properties of their parent transforms. For example, linear wavelet transforms are known to be extremely effective for decorrelation and also have useful multiresolution properties. For all of the reasons described above, reversible ITI wavelet transforms are an extremely useful tool for sig-nal coding applications. Such transforms can be employed in lossless coding systems, hybrid lossy/lossless coding systems, and even strictly lossy coding systems as well. 1.3 Common Misconceptions Before proceeding further, it might be wise to debunk a number of common myths about reversible ITI wavelet transforms. In so doing, we may avoid the potential for confusion, facilitating easier understanding of subsequent material in this thesis. Although reversible ITI wavelet transforms map integers to integers, such transforms are not fundamentally integer in nature. That is, these transforms are based on arithmetic over the real numbers in conjunction with rounding operations. Unfortunately, this is not always understood to be the case. The frequent abuse of terminology in the research literature is partially to blame for the perpetuation of this myth. Many authors abbreviate the qualifier "integer-to-integer" simply as "integer". This, however, is a confusing practice, as it actually suggests that the transforms are based solely on integer arithmetic (which is not the case). In the interest of technical accuracy, this thesis will always use the qualifier "integer-to-integer". When real arithmetic is implemented on a computer system, a (finite- and) fixed-size word is typically employed for representing real quantities. This, however, has the implication that only a finite number of real values can be represented exactly. Consequently, some arithmetic operations can yield a result that cannot be represented exactly, and such a result must be rounded to a machine-representable value. For this reason, real arithmetic is not normally exact in a practical computing system. Two different types of representations for real numbers are commonly employed: fixed point and floating point. If the radix point is fixed at a particular position in the number, the representation is said to be fixed point. Otherwise, the representation is said to be floating point. Fixed-point arithmetic is typically implemented using machine instructions operating on integer types, and is much less complex than floating-point arithmetic. It is important to understand the above concepts and definitions in order to debunk the next myth about reversible ITI wavelet transforms. (For a more detailed treatment of computer arithmetic, we refer the reader to [42].) One must be mindful not to make the mistake of confusing the theoretical definition of a transform with its practical implementation. Unfortunately, this happens all too often. For example, reversible ITI transforms are often (incor-rectly) extolled as being superior to conventional (i.e., real-to-real) wavelet transforms in a computational complexity sense, simply because the former can be computed using fixed-point arithmetic. Such a belief, however, is misguided, since both reversible ITI and conventional wavelet transforms can be implemented using fixed-point arithmetic. For that matter, both types of transforms can also be implemented using floating-point arithmetic (although, in practice, one would probably never realize reversible ITI transforms in this manner). Although reversible ITI wavelet trans-forms do (potentially) have computational/memory complexity advantages over their conventional counterparts, these advantages cannot be attributed solely to the use of fixed-point arithmetic, as explained above. We will have more to say on this topic later in the thesis. 1.4 Historical Perspective In some sense, reversible ITI wavelet transforms are not new. The S transform [28,1.08], a reversible ITI version of the Haar wavelet transform [62], has been known for at least a decade. What has not been clear until more recently, Chapter 1. Introduction 3 however, was how the ideas behind the S transform could be extended in order to obtain a general framework for reversible ITI wavelet transforms. As a consequence, the S transform remained, for a number of years, the most effective reversible ITI wavelet transform known. General frameworks for reversible ITI wavelet transforms are a more recent innovation. In fact, many of the key developments associated with such frameworks have occurred only in the last several years. The common element in all of the frameworks proposed to date is the ladder network. Wavelet transforms are realized using uniformly maximally-decimated (UMD) filter banks, and the UMD filter banks are implemented in polyphase form using ladder networks to perform the polyphase filtering. Such networks are of critical importance as they can be made invertible even in the presence of various types of quantization error, especially the rounding error introduced by finite-precision arithmetic. In 1992, Breukers and van den Enden [36] were the first to propose invertible polyphase networks based on ladder structures. Soon afterward, such networks were studied in detail by Kalker and Shah [81,82]. A few years later, more examples of reversible ITI transforms began to appear in the literature (e.g., [ 148,87]). In fact, Komatsu et al. [87] even proposed a transform based on nonseparable sampling/filtering. Around this time, Said and Pearlman [114] proposed the S+P family of transforms, an extension of the S transform which made use of ladder structures. Despite of all of this work, however, the full potential of ladder networks had not yet been fully appreciated. In 1995, Sweldens [ 1.29,130] initially proposed the lifting framework for the design and implementation of wavelet transforms—this new framework being based on ladder-type polyphase networks. At this time, he first coined the term "lifting", and detailed some of the many advantages of this realization strategy. Soon afterward, Daubechies and Sweldens [49] demonstrated that any (appropriately normalized) one-dimensional (ID) two-band biorthogonal wavelet transform (associated with finitely-supported wavelet and scaling functions) can be realized using the lifting framework. This development further increased the popularity of lifting. Sweldens later authored several other papers related to lifting including [131]. As it turns out, determining the lifting realization of a wavelet transform is equivalent to a matrix factorization problem. More specifically, one must decompose the polyphase matrix of a UMD filter bank into factors of a particular form. In the general D-dimensional M-band case, a solution to this problem may not exist. In some cases, however, a solution must exist as a consequence of a mathematical result known as Suslin's stability theorem [128]. Recently, a constructive proof of this theorem was devised by Park and Woodburn [106]. Tolhuizen et al. [137] have also devised a related factorization algorithm. These works are practically significant as they provide a means by which to perform the factorization (in the general case), and thus find a lifting realization of a transform (when such exists). Although ladder-based polyphase networks, such as those employed by the lifting framework, are ideally suited for the construction of reversible ITI wavelet/subband transforms, this fact was not realized until 1996. At this time, Calderbank et al. [38] (later published as [40]) first demonstrated that the lifting scheme forms an effective framework for constructing reversible ITI approximations of linear wavelet transforms. Furthermore, they showed that such trans-forms could be used to good effect for lossless signal coding. This work was of crucial importance, as it established a linkage between all of the previous studies of ladder-based polyphase networks and reversible ITI wavelet transforms. Also, around the same time as the discovery by Calderbank et al., several other researchers proposed some similar ideas (e.g., Chao et al. [44] and Dewitte and Cornelis [50]). More recently, Jung and Prost [79] have proposed an alternative method for constructing reversible ITI transforms based on the overlapping rounding transform (ORT)—the ORT again being based on ladder networks. Also, L i et al. [93] have considered the more general problem of creating reversible ITI versions of arbitrary linear transforms. In the context of wavelet/subband transforms, their work essentially suggests the use of ladder networks with IIR filters (in addition to FIR filters). So far, in this historical commentary, we have focused our attention primarily on the events leading to the de-velopment of general frameworks for reversible ITI wavelet transforms. Of course, the frameworks themselves are not particularly useful unless methods exist for designing effective transforms based on these frameworks. As sug-gested earlier, this design problem is one of perfect-reconstruction (PR) UMD filter bank design. In the ID case, the PR UMD filter bank design problem has been extensively studied and a multitude of design techniques have been proposed (e.g., [140,125,141]). Although the multidimensional case has received much less attention, numerous tech-niques have been devised (e.g., [90,98,95,117, .136,45,105,48]). Unfortunately, very few of the design techniques proposed to date specifically consider UMD filter banks with the reversible and ITI properties. For this reason, many of the previously proposed design methods are either not directly useful or impractical to apply. Of the PR UMD filter bank design methods proposed to date, the most useful ones are those based on ladder-type polyphase networks (e.g., lifting). Recent work by Kovacevic and Sweldens [90] and Marshall [98] both consider such Chapter 1. Introduction 4 networks. Other work by Redmill [II1] has examined the ID case. Alternatively, one can design PR UMD filter banks based on arbitrary filtering structures and then attempt to find an equivalent ladder-based realization. In this regard, the results of [49, 106, 1.37] are helpful. Such an approach may not be the most attractive, however. Both nonexistence and nonuniqueness of the ladder-based realization pose potential problems. 1.5 Overview and Contribution of the Thesis This thesis studies reversible ITI wavelet transforms and their application to image coding. In short, this work examines such matters as: transform frameworks, rounding operators and their characteristics, transform design techniques, strategies for handling the transformation of arbitrary-length signals in a nonexpansive manner, computational and memory complexity issues for transforms, and the utility of various transforms for image coding. Structurally, this thesis is organized into six chapters and one appendix. The first two chapters provide the back-ground information necessary to place this work in context and facilitate the understanding of the research results presented herein. The remaining four chapters present a mix of research results and additional concepts required for the comprehension of these results. The appendix provides supplemental information about related topics of interest, but such details are not strictly necessary for an understanding of the main thesis content. Chapter 2 begins by presenting the notation, terminology, and basic concepts essential to the understanding of this thesis. Multirate systems are introduced, and UMD filter banks are examined in some detail. Then, the link between UMD filter banks and wavelet systems is established. Following this, a brief introduction to image coding is given. Chapter 3 examines frameworks for reversible ITI wavelet/block transforms. Ideas from previously proposed frameworks are integrated and extended in order to construct a single unified framework known as the generalized reversible ITI transform (GRITIT) framework. This new framework is then used to study the relationships between the various other frameworks. For example, we show that the framework based on the overlapping rounding transform (ORT) is simply a special case of the lifting framework with only trivial extensions. Next, the GRITIT framework is examined in the context of block transforms. We propose the generalized S trans-form (GST), a family of reversible ITI block transforms. This family of transforms is then studied, by considering top-ics such as the relationship between the GST and the GRITIT framework, GST parameter calculation, and the effects of using different rounding operators in the GST. We show that, for a fixed choice of integer-bias invariant rounding operator, all GST-based approximations to a given linear transform are equivalent. The reversible color transform (RCT) is shown to belong to the GST family, and a new transform in this family, called the modified reversible color transform (MRCT), is also proposed and shown to be effective for image coding purposes. We demonstrate how the various degrees of freedom in the choice of GST parameters can be exploited in order to find transform implementa-tions with reduced computational complexity. For example, the GST framework is used to derive lower complexity implementations of the S transform and RCT. In this chapter, we also focus our attention on rounding operators. Such operators may potentially be integer-bias invariant or odd. These two attributes are shown to be mutually exclusive, and their significance is also discussed. Several common rounding operators are considered. Relationships between these operators are examined, and these operators are also characterized in terms of their approximation properties (i.e., error interval, peak absolute error, and mean absolute error) and computational complexity. In so doing, we show that the floor operator incurs the least com-putation, while, under most circumstances, the biased floor and biased ceiling operators share the best rounding error performance at a slightly higher computational cost. We also demonstrate that, in some cases, the floor operator can actually share the best error performance (in spite of its low computational complexity), a fact that is often overlooked in the existing literature. Chapter 4 examines various strategies for handling the transformation of arbitrary-length signals in a nonexpan-sive manner, including symmetric extension and per-displace-step (PDS) extension. In this context, two families of symmetry-preserving transforms are introduced, and the transforms from both families are shown to be compatible with symmetric extension, which allows arbitrary-length signals to be treated in a nonexpansive manner. The charac-teristics of the two families and their constituent transforms are then studied. We identify which transforms belong to each family. For the more constrained of the two families, we show that 1) such transforms are associated with analysis filters having transfer functions of a highly structured form, 2) the DC and Nyquist gains of these analysis filters are fixed, independent of the choice of free parameters, and 3) if a particular filter bank is associated with a transform in this family, then so too is its "transposed" version. Some new reversible ITI structures that facilitate the construction Chapter 1. Introduction 5 of symmetry-preserving transforms are also proposed. In addition, a simple exhaustive search technique is explored as a means for finding good low-complexity transforms in the above-mentioned families. Several new transforms are found with this approach and shown to be effective for image coding. Next, the relationship between symmetric extension and PDS extension is studied. We show that, in some specific cases, symmetric extension is equivalent to constant PDS extension. We explain how this equivalence can be exploited in order to simplify JPEG-2000 codec implementations. Chapter 5 studies the use of reversible ITI wavelet transforms for image coding. In this context, a number of reversible ITI wavelet transforms are compared on the basis of their lossy compression performance, lossless com-pression performance, and computational complexity. Of the transforms considered, several were found to perform particularly well, with the best choice for a given application depending on the relative importance of the preceding criteria. Transform-coefficient dynamic range is also examined and its implications on word size and memory com-plexity considered. Reversible ITI versions of numerous transforms are also compared to their conventional (i.e., non-reversible real-to-real) counterparts for lossy compression. At low bit rates, reversible ITI and conventional versions of transforms were found to often yield results of comparable quality. Factors affecting the compression performance of reversible ITI wavelet transforms are also presented, supported by both experimental data and theoretical arguments. Finally, Chapter 6 summarizes some of the more important results presented in this thesis along with the contribu-tions that it makes. The chapter concludes by suggesting directions for future research. Much of the work presented herein was conducted in the context of JPEG-2000 standardization activities. In fact, many of the image coding results were obtained with various implementations of the JPEG-2000 codec (at varying stages of development). For this reason, the JPEG-2000 standard is of direct relevance to this thesis, and a detailed tutorial on the standard is provided in Appendix A. One of the JPEG-2000 codec implementations used to obtain some of the experimental results presented herein belongs to the JasPer software. A very brief description of this software can also be found in the same appendix. Those who educate children well are more to be honored than parents, for these only gave life, those the art of living well. —Aristotle Chapter 2 Preliminaries 6 There is only one thing you should do. Go into yourself. Find out the reason that commands you to write; see whether it has spread its roots into the very depths of your heart; confess to yourself whether you would have to die if you were forbidden to write. This most of all: ask yourself in the most silent hour of your night: must I write? Dig into yourself for a deep answer. And if this answer rings out in assent, if you meet this solemn question with a strong, simple "I must," then build your life in accordance with this necessity; your whole life, even into its humblest and most indifferent hour, must become a sign and witness to this impulse. —Rainer Maria Rilke, Letters to a Young Poet (First letter, 1903) 2.1 Introduction In order to facilitate the understanding of the research results presented in this thesis, we must first explain some of the fundamental concepts relevant to this work. Such preliminaries are provided by this chapter. First, we briefly introduce some of the basic notation and terminology used herein. Then, we proceed to give a tutorial on multirate filter banks and wavelet systems. Lastly, we introduce some basic concepts from image coding. 2.2 Notation and Terminology Before proceeding further, a brief digression concerning the notation used in this thesis is appropriate. In the remainder of this section we introduce some of the basic notational conventions and terminology employed herein. The symbols Z, R, and C are used to denote the sets of integer, real, and complex numbers, respectively. The symbol j denotes the quantity y/—i. Matrix and vector quantities are indicated using bold type (usually in upper and lower case, respectively). The transpose and transposed inverse of a matrix A are denoted, respectively, as AT and A~T. The symbols IN and JN denote theNxN identity and anti-identity matrices, respectively, where the subscript N may be omitted when clear from the context. A matrix A is said to be unimodular if |det A\ = 1. The (i,y)th minor of the N x N matrix A, denoted minor(>l, is the (N-l)x(N- 1) matrix formed by removing the rth row and y'th column from A. In the case of matrix multiplication, we define the product notation as follows: I7JLM A; = ANAN-\ • • • AM+\ AM for N >M. One should note the order in which the matrix factors are multiplied above, since matrix multiplication is not commutative. In the case that N < M, we define the product notation td denote an "empty" product (i.e., the multiplicative identity, JT). The notation Pr(x) denotes the probability of event x, and the notation E(JC) denotes the expected value of the quantity x. For two vectors z = [zo zi - zD-\]T and fc = [*o h - fcD_i]r, we define the scalar quantity Chapter 2. Preliminaries 1 Moreover, for a D x D matrix M, we define Z M = [zm0 ... z m D - l ] T , where is the M i column of M. For a Laurent polynomial P(z), we denote the degree of P(z) as degZ'(z), and define this quantity as follows. In the case that P(z) has the form P(z) = 1-LMpiz~l where N > M , p M ^ 0 , and p N ^ 0 (i.e., />(z) ^ 0), deg/>(z) = N - M . In the case that P(z) is the zero polynomial (i.e., P(z) = 0), we define deg/>(z) = Thus, for any two Laurent polynomials, A(z) and B(z), we have that deg (A(z)S(z)) = degA(z) + degfl(z). For a £ K, the notation |_aj denotes the largest integer not more than a (i.e., the floor function), and the notation [a] denotes the smallest integer not less than a (i.e., the ceiling Junction). One can show that the following relationships hold for the floor and ceiling functions: [x+cxj =x+[a\, (2.1) \x + a\ = x + \a], and (2.2) fal = - L - a J , (2.3) for all a £ R and all x g Z . The biased floor, biased ceiling, truncation, biased truncation, and rounding-away-from-zero (RAFZ) functions are defined, respectively, as bfloora = [a + £ J , (2.4) bceila= [ a - i ] , (2.5) A f |aj for a > 0 trunca= < Lr i (2.6) [ fal for a < 0, A (bfloora fora>0 btrunca=< ~ (2.7) I bceila for a < 0, and A f Tal for a > 0 rafza = { \ ~ (2.8) for a < 0, where a € IR. The mapping associated with the biased truncation function is essentially equivalent to traditional rounding (to the nearest integer). The signum function is defined as sgna = < 1 for a > 0 0 for a = 0 -1 fora<0, where a e K. The fractional part of a real number a is denoted fraca and defined as fraca = a - [aj. Thus, we have that 0 < fraca < 1 for all a e R. We define the mod function as mod(jt,y) = x — y [x/y\ where x , y £ Z . (2.9) In passing, we note that the mod function simply computes the remainder when x is divided by y, with division being defined in such a way that the remainder is always nonnegative. From (2.9) and (2 .3 ) , we trivially have the identities [x/y\ = *-™^(*<y) fory^O, and (2.10) \x/y\ = fory^O. (2.11) Suppose that Q denotes a rounding operator. In this thesis, such operators are defined only in terms of a single scalar operand. As a notational convenience, however, we use an expression of the form Q(x), where a; is a vector/matrix Chapter 2. Preliminaries 8 quantity, to denote a vector/matrix for which each element has had the operator Q applied to it. A rounding operator Q is said to be integer invariant if it satisfies Q { x ) = x for all x € Z (2.12) (i.e., Q leaves integers unchanged). For obvious reasons, rounding operators that are not integer invariant are of little practical value. Thus, we consider only integer-invariant rounding operators in this work. A rounding operator Q is said to be integer-bias invariant if Q ( a + x ) = Q { a ) + x for all a e R and a l lx e Z. (2.13) Similarly, a rounding operator Q is said to be odd if Q(a) = -e(-oc) for all a e i . (2.14) We will later show that a rounding operator cannot be both odd and integer-bias invariant. In passing, we note that the floor, biased floor, ceiling, and biased ceiling functions are integer-bias invariant (but not odd), while the truncation, biased truncation, and RAFZ functions are odd (but not integer-bias invariant). All rounding operators considered in this thesis are tacitly assumed to be memoryless and shift invariant. Any (reasonable) rounding operator will preserve signal symmetry (but not necessarily signal antisymmetry), while any odd rounding operator will preserve both signal symmetry and antisymmetry. The z-transform of a sequence x[n] is denoted as Zx[n] (OTX(Z)) and is defined as Zx[n]=X(z)= X x[n]z~n. nSZ° The inverse z-transform operator is denoted as Z r x . The notation f * g stands for the convolution of / and g. For convenience, we also define the quantities [ X ( z ) J J = 2 > M J z - n and n€Z [X(z)i ,= 2r*Miz-" n€Z (i.e., [x[n}\ <-> [ X ( z ) \ z ^ d \x[n}] <- \ X ( z ) ] z ) . Also, the floor and ceiling operators for sequences are shift invariant. That is, we have that [ z K X ( z ) \ z = z K [ X ( z ) \ z and \ z K X ( z ) \ = z K \ X ( z ) \ where K e Z. A (linear) D-point block transform is a transformation that maps an input vector x = [*o *i - *D-\] to the output vector y = [yo y\ - yD-i ] , and can be expressed in the form y = A x where A is a D x D matrix. The discrete Fourier transform (DFT) and discrete cosine transform (DCT) are examples of commonly used block transforms. 2.3 Multirate Filter Banks and Wavelet Systems Multirate filter banks play an important role in the study of wavelet systems. In particular, the class of systems known as uniformly maximally-decimated (UMD) filter banks is especially significant in this regard. Wavelet transforms can be constructed through the design of UMD filter banks and implemented very efficiently in terms of these structures. In order to study UMD filter banks, we must first understand the fundamentals of multirate systems. In what follows, we begin by introducing some basic multirate system concepts, and then use these concepts to establish a Chapter 2. Preliminaries 9 general framework for the study of UMD filter banks. Finally, the link between UMD filter banks and wavelet systems is established. The theory of multirate filter banks and wavelet systems is quite involved, and one could easily devote several books to this topic. Consequently, due to space constraints, we cannot hope to provide a comprehensive introduction to this theory here. We will only briefly present some of the fundamental concepts that are most relevant to our work. For information beyond that which we present here, one might consider further reading in the following areas: mul-tirate systems and filter banks [140,141,125,139,37], multidimensional signal processing [53,94], multidimensional multirate systems [138,143,83], multidimensional sampling [52], and multidimensional wavelet theory [91.]. 2.3.1 Multirate Systems A system is said to be multirate if signals at different points in the system are defined on different sampling lattices. For the purpose of this discussion, a lattice is simply a regular periodic configuration of points in some space. The physical significance of the lattice axes depends on the particular system involved. For example, in the case of a system that processes images (which are 2D signals), the lattice axes correspond to two directions of displacement in physical space. In the case of a system that processes audio signals (which are ID), the single lattice axis corresponds to time, and a multirate system would be one that employs more than one sampling rate. 2.3.2 Sampling In multirate systems, we obviously need a means for identifying how signals are sampled. In this context, some basic definitions and other concepts from lattice theory are of great importance. As suggested previously, a lattice is a regular periodic configuration of points in some space. Often, we are also interested in a sublattice, which is a lattice comprised of a subset of the points from another lattice. Of particular interest to us here is the integer lattice. The D-dimensional integer lattice is simply the set of all D-dimensional integer vectors (i.e., all vectors of the form n e 1 D ) . Sublattices of the integer lattice can be specified using a sampling matrix. Such a matrix is square, nonsingular, and has only integer entries. The sublattice (of the integer lattice) generated by the D x D sampling matrix M is denoted as LAT(M), and is defined as the set of all vectors of the form M n where n is a D-dimensional integer column vector. In other words, L A J ( M ) is the set of all integer linear combinations of the columns of M . If a sampling matrix is diagonal, then the corresponding sampling scheme is said to be separable; otherwise, the scheme is said to be nonseparable. In general, the sampling matrix generating a particular lattice is not unique. In fact, for any sampling matrix M and unimodular integer matrix U , one can show [52] that LAT(M) and L A T ( M U ) (i.e., the sublattices generated by M and M U ) are the same. Some examples of lattices are illustrated in Figure 2.1. The rectangular and quincunx lattices shown in the figure can be generated, respectively, with the following sampling matrices: and M i = ^ j . A coset of a sublattice is the set of points obtained by shifting the entire sublattice by an integer vector. Moreover, a shift vector associated with a particular coset is referred to as a coset vector. The number of distinct cosets of LAT(iVf) is |detM|, and the union of these cosets is the integer lattice. The quantity |detM| frequently appears in mathematical expressions and physically represents the reciprocal of the sampling density. In other words, LAT(M) (i.e., the lattice generated by M ) has J^J^J the sampling density of the integer lattice. 2.3.3 Downsampling Downsampling is one of the fundamental processes in multirate systems, and is performed by a processing element known as the downsampler. The downsampler, shown in Figure 2.2, takes an input signal x[n] and produces the output signal y ( n ) = x { M n ) (2.16) where M is a sampling matrix. In simple terms, this process samples the input x \ n \ by mapping the points on the sublattice LAT(M) to the integer lattice and discarding all others. The relationship between the input and output of M 0 -2 0 0 2 Chapter 2. Preliminaries 10 • 5-• 4-• 3-•2H \ \ I ) t I T t t I ••• 1 2 3 4 5 n0 n\. 5 — 4--3"-2 " 1--+—f—I ( 1 ••• 1 2 3 4 5 no "1: ' 5-4 + 3" 2 + 1-I t I 1 2 3 5 no (a) (b) (c) Figure 2 . 1 : Examples of 2D lattices, (a) The integer lattice, (b) A rectangular lattice, (c) The quincunx lattice. y[n] Figure 2.2: Downsampler. the downsampler in the z-domain is given by M-I ;=o where .r™„ _;,. ,r r a . = ^ S X ( e M - 1 ( 2 7 i f c , ) z M " ) , (2.17) M = | d e t M | , e M (w) and rrifc is the ^th column of M. This leads directly to the frequency domain relation i M-l = M E X ^ (M- 7"(w-2nfc,)). (2.18) M (=0 In the ID case, the quantities n, z, and u; can be replaced with their respective scalar equivalents n, z, and co, and M becomes synonymous with the scalar quantity M. Thus, in this special case, we can simplify (2.16), (2.17), and (2.18), respectively, to obtain y(n)=x{Mn), (2.19a) i M-l Y{z) = — X X(e-2MMzl,M), and (2.19b) M i=0 i M - l yF(FFL) = 1 7 S ((CO-2JU-)/W). (2.19C) M i=0 From (2.18) and (2.19c), we can see that downsampling has a simple interpretation in the frequency domain. That is, the spectrum of a downsampled signal is merely the average of M shifted and "stretched" versions of the original input spectrum. Evidently, downsampling can result in multiple baseband frequencies in the input signal being mapped to a single frequency in the output signal. This phenomenon is called aliasing. If aliasing occurs, it is not possible to recover the original signal from its downsampled version. Chapter 2. Preliminaries 11 y[n] Figure 2.3: Upsampler. 2.3.4 Upsampling Upsampling is another fundamental process in multirate systems, and is performed by a processing element known as the upsampler. The upsampler, shown in Figure 2.3, takes an input signal x[n] and produces the output signal f i f M - ' n ) i f n S L A T ( M ) ( 2 2 Q ) 10 otherwise where M is a sampling matrix. One can see that this process maps a signal on the integer lattice to another signal that is nonzero only at points on the sampling sublattice LAT(M) . In the z-domain, the relationship between the input and output of the upsampler is given by Y { z ) = X ( z M ) . (2.21) This directly yields the frequency domain relation Y F ( u ) = X F ( M T u > ) . (2.22) In the ID case, the quantities n , z, and u> can be replaced with their respective scalar equivalents n, z, and oo, and M becomes synonymous with the scalar quantity M . Thus, in this special case, we can simplify (2.20), (2.21), and (2.22), respectively, to obtain y [ n ] = x [ M n ] , (2.23a) Y ( z ) = X ( z M ) , and (2.23b) YF((o) = Xf(M<O). (2.23C) From (2.22) and (2.23c), we can see that upsampling has a simple interpretation in the frequency domain. That is, the spectrum of the upsampled signal is just a "compressed" version of the input signal spectrum. In general, we obtain exactly M complete images of one period of X f ( u ) in />(>>). This phenomenon is called imaging. The original signal can be recovered from its upsampled version by removing the M — 1 superfluous images of the original signal baseband. 2.3.5 Noble Identities Often an upsampler or downsampler appears in cascade with a filter. Although it is not always possible to interchange the order of upsampling/downsampling and filtering without changing system behavior, it is sometimes possible to find an equivalent system with the order of these operations reversed, through the use of two very important relationships called the noble identities. The first identity, illustrated in Figure 2.4(a), allows us to replace a filtering operation on one side of a downsampler with an equivalent filtering operation on the other side of the downsampler. The second identity, shown in Figure 2.4(b), allows us to replace a filtering operation on one side of an upsampler with an equivalent filtering operation on the other side of the upsampler. In addition to their theoretical utility, the noble identities are of great practical significance. For performance reasons, it is usually desirable to perform filtering operations on the side of an upsampler or downsampler with the lower sampling density. Using the noble identities, we can move filtering operations across upsamplers/downsamplers in order to achieve improved computational efficiency. 2.3.6 Polyphase Representation of Signals and Filters A fundamental tool in the study of multirate systems is the polyphase representation of a signal. Such a representation is defined with respect to a particular sampling matrix and corresponding set of coset vectors, and serves to decompose Chapter 2. Preliminaries 12 x[n] | M r G(z) y[n] : x[n] G{zM) y[n] U(z) G(z) (a) y[n] x[n] (b) T M G{zM) y[n] Figure 2.4: Noble identities. The (a) first and (b) second identities. H0(zM) (b) U(z) H(z) V(z) (a) ~ V(x) U(z) W l ( 2 M ) i Wo(2 M ) (c) H,{zM) i Figure 2.5: Polyphase form of a filter, (a) Original filter. The (b) type-1 and (c) type-2 polyphase representations of the filter. a signal into a sum of signals called polyphase components. Two types of polyphase representations are commonly used, which we refer to as type 1 and type 2. Mathematically, the polyphase representation of the signal x[n], with respect to the sampling matrix M and corresponding set of coset vectors {mi}fs0x, is defined as M - l x[n] = X * / M (=0 (2.24) where xi[n] \x[Mn + mi] type 1 [ j [ M n - mi] type 2, and M = |detM|. In the above equation, xt [n] denotes the Ith polyphase component of x[n). In the z-domain, we have 1 ^hlo'z^'MzM) type 2. Not only can we represent signals in polyphase form, but we can also represent filters in this manner. In this case, the signal x[n] that we decompose in (2.24) is the impulse response of the filter in question. This process yields a number of polyphase filters (which correspond to the */[rc]). This leads to the structures shown in Figure 2.5 for the type-1 and type-2 polyphase representations of a filter. As we shall see, the polyphase representation is often a mathematically convenient form in which to express filtering operations in multirate systems, facilitating easier analysis of such systems and greatly simplifying many theoretical results. Also, this representation leads to an efficient means for implementing filtering operations in a multirate framework, as elaborated upon below. In multirate systems, filters are often connected in cascade with upsamplers/downsamplers. For reasons of compu-tational efficiency, it is usually preferable to perform any filtering on the side of the upsampler/downsampler with the Chapter 2. Preliminaries 13 x[n] Ho(z) vol"] v i H Fo(z) Fi(z) HM-i{z) uM-\[n\ VM-\\n\ FM-I(Z) x[n] Figure 2.6: Analysis bank. Figure 2.7: Synthesis bank. lower sampling density. We have seen how the noble identities can be used to move a filtering operation across upsam-plers/downsamplers, but in order to move filtering to the side of the upsampler/downsampler with the lower sampling density, the transfer function of the filter must be an expression in zM which is not generally the case. The polyphase representation, however, provides us with a means to express any filter as a set of filters with transfer functions in z M so that filtering can be moved to the more desirable side of an upsampler or downsampler. 2.3.7 Filter Banks A filter bank is a collection of filters having either a common input or common output. When the filters share a common input, they form what is called an analysis bank. When they share a common output, they form a synthesis bank. These two types of filter banks are depicted in Figures 2.6 and 2.7. Each of the filter banks shown consists of M filters. The filters {H*} belonging to the analysis bank are called analysis filters and the filters {F^} comprising the synthesis bank are referred to as synthesis filters. The signals {w/t[n]} and {vjt[n]} are called subband signals. The frequency responses of the analysis/synthesis filters may be non-overlapping, marginally overlapping, or greatly overlapping depending on the application. 2.3.8 Uniformly Maximally-Decimated (UMD) Filter Banks Although many filter bank configurations exist, an extremely useful one is the so called uniformly maximally-decimated (UMD) filter bank1. The general structure of such a system is shown in Figure 2.8 where M = |detM|. The analysis bank and downsamplers collectively form the analysis side of the UMD filter bank. Similarly, the upsamplers and synthesis bank together constitute the synthesis side of the UMD filter bank. The UMD filter bank (shown in Figure 2.8) operates in the following manner. On the analysis side, the input signal x[n] is processed by the analysis filters {H*}, and the resulting filter outputs are then downsampled by M , yielding the subband signals {yjt[n]}. On the synthesis side, the subband signals {>»*;[«•]} are upsampled by M . Then, the upsampler outputs are transformed by the synthesis filters {F^}, and the resulting filter outputs are summed to produce the output signal x [ n ] . Each of the M subband signals {yjt[n]} has ^-th the sampling density of the original input signal x[n]. Conse-quently, the subband signals collectively possess the same number of samples as the original input signal. For this reason, the filter bank is referred to as maximally decimated. Furthermore, since all of the subband signals have the same sampling density, the filter bank is said to be uniformly decimated. Hence, a filter bank of the above form is referred to as uniformly maximally decimated. 2.3.9 Perfect-Reconstruction (PR) UMD Filter Banks For the UMD filter bank of Figure 2.8 to be of practical use, the output x[n] is usually required to be an accurate reproduction of the input x [ n ] . If the system is such that x[n] = x [ n — no] for all x[n] and for some integer vector no, the system is said to have the perfect reconstruction (PR) property. In other words, a PR system can reproduce the 'Sometimes, the qualifier "quadrature mirror-image" (abbreviated QM) is used as a synonym for "uniformly maximally decimated", but strictly speaking this is an abuse of terminology. Chapter 2. Preliminaries 14 T , IM yM-l[n\ T M , I Analysis Bank Downsamplers Upsamplers Synthesis Bank Analysis Side Synthesis Side Figure 2.8: Canonical form of a D-dimensional Af-channel UMD filter bank. input signal exactly except for a possible shift. In the special case that no = 0, the system is said to possess the shift-free PR property. Generally, there are three reasons that the reconstructed signal x[n] can differ from x[n]: aliasing distortion, amplitude distortion, and phase distortion. The analysis and synthesis filters can be designed in such a way so as to eliminate some or all of these distortions depending on what is required by the application. (In practice, the effects of finite-precision arithmetic can also introduce distortion, but we are only considering the theoretical reasons for distortion here.) Since our interests lie with invertible transforms, we focus exclusively on UMD filter banks having the PR property in this thesis. 2.3.10 Polyphase Form of a UMD Filter Bank Although the structure for the UMD filter bank shown in Figure 2.8 may be intuitively appealing, it is often not the most convenient structure with which to work. This leads us to the polyphase form of a UMD filter bank, which is related to the polyphase representation of individual filters (as discussed earlier). The polyphase representation has many advantages, but most importantly it simplifies many theoretical results and suggests^ an efficient means of implementation. ^ Suppose that we have a UMD filter bank with analysis filters {H*}, synthesis filters {F^}, sampling matrix M, and the corresponding coset vectors {m/J-Jijj1, where M = |detM|. For notational convenience, we define the quantities h(z) ' H0(z) " " F0(z) • • z m 0 ' , u(z) = , /(*) = HM-\{Z)_ z™M-l and v(z) •• ,-mo ,-m\ - « i M - l First, let us consider the analysis filters of the UMD filter bank. We can express the transfer functions in type-2 polyphase form as M - l Hk(z) = X z ^ E u ( z M ) . 1=0 This equation can be rewritten in matrix form as E0fi(zM) £ i , o ( z M ) " H0(z) ' Hi(z) = HM-\(Z)_ Eo,l(zM) Ei,i(zM) £O,M-I (Z^) 1 R EI,M-I(Z ) or more compactly as h{z) = E(zM)u(z) (2.25) Chapter 2. Preliminaries 15 E{zM) | M | M T M J M T M R{zM) Figure 2.9: Polyphase realization of a D-dimensional M-channel UMD filter bank before simplification using the noble identities. where E(z) = E0,o(z) £b,i(«) £1,0(2) ElA(z) \_EM-I,O(Z) EM-iti(z) ••• £M-I,A/-I(^)J Equation (2.25) completely characterizes the analysis bank and is called the polyphase representation of the analysis bank. The quantity E[z) is referred to as the analysis polyphase matrix. Now, let us consider the synthesis filters of the UMD filter bank. We can express the transfer functions (F*(z)} in type-1 polyphase form as A/-1 Fk(z) = X z-m'Rltk(zM). 1=0 In matrix form, this becomes [F0(z) F,(z) ••• F„-X(z)] = Ro,o(zM) *o, i (z M ) Rho(zM) Rx,x(zM) | R M - I , O ( Z M ) RM-XAZM) RO,M-I(Z™) Rl,M-l(z ) RM-\,M-I{Z ) which can be written more concisely as fT(z) = vT(z)R(zM) (2.26) where R(z) = Ro,o(z) *i,o(«) *b,i(*) RiAz) Ro,M-l{z) R\,M-\(Z) [_RM-\,O(Z) RM-I,I{Z) ••• RM-\,M-\{Z)\ Equation (2.26) completely characterizes the synthesis bank and is called the polyphase representation of the synthesis bank. The quantity R(z) is referred to as the synthesis polyphase matrix. Observe that (2.25) and (2.26) provide an alternative way in which to express the analysis and synthesis banks of the UMD filter bank. Suppose now that we have a UMD filter bank where the analysis and synthesis banks have been represented in this manner. In this case, these equations give us the transformed, but mathematically equivalent, system shown in Figure 2.9. Using the noble identities, however, we can move the analysis polyphase filtering to the right of the downsamplers and the synthesis polyphase filtering to the left of the upsamplers. This gives us the polyphase form of the filter bank shown in Figure 2.10. In effect, the polyphase representation reorganizes a filter bank so that it operates on the polyphase components of the input signal. The analysis and synthesis polyphase filtering is performed by M-input M-output networks. On the analysis side of the filter bank, the shift operators and downsamplers form what is referred to as the forward polyphase Chapter 2. Preliminaries 16 *[n] 1*4 | M yo[n\ yiH » - i H TMh TM 2 - m 0 Figure 2.10: Polyphase realization of a D-dimensional M-channel UMD filter bank. transform (FPT). Similarly, the upsamplers and shift operators form what is called the inverse polyphase transform (WT). It is important to note that with the polyphase form of a UMD filter bank, analysis filtering is performed after downsampling, and synthesis filtering is performed before upsampling. In other words, all filtering is performed in the downsampled domain (i.e., at the lower sampling density). This leads to improved computational efficiency over a UMD filter bank implemented in its canonical form. Consequently, most practical implementations of UMD filter banks make use of the polyphase form. Moreover, the polyphase form is of great importance in the context of reversible ITI transforms, as will become apparent later. 2.3.11 Conditions for PR System Many applications necessitate the use of a filter bank with the PR property. Moreover, we often further require that a filter bank possess the shift-free PR property. Therefore, it is only natural to wonder what conditions a filter bank must satisfy in order to have this property. One can show that a UMD filter bank has the shift-free PR property if and only if R(z)E(z) = I (2.27) where E(z) and R(z), respectively, denote the analysis and synthesis polyphase matrices of the filter bank. By examining the polyphase form of a UMD filter bank (shown in Figure 2.10), we can see the reason behind the above condition for shift-free PR. If (2.27) is satisfied, then the synthesis polyphase filtering (represented by R(z)) cancels the effects of the analysis polyphase filtering (represented by E(z)). This being the case, the filter bank, in effect, only serves to split a signal into its polyphase components and then recombine these components, yielding the original input signal with no shift. . 2.3.12 Octave-Band Filter Banks As shown in Figure 2.11, the analysis side of a UMD filter bank decomposes the input signal x[n] into M subband signals {^[n]}. The synthesis side then recombines these subband signals to obtain x[n], the reconstructed version of the original signal. There is nothing, however, to prevent the use of additional UMD filter banks to further decompose some or all of the subband signals {yt[n]}. Of course, some or all of the resulting subband signals can again be decomposed with even more UMD filter banks. In other words, this idea can be applied recursively, and the final result is a filter bank with a tree structure. If a tree-structured UMD filter bank is such that 1) only the lowpass subband signal is decomposed at each level in the tree, 2) the same basic UMD filter bank building block is used for decomposition at all levels, and 3) this basic block has PR and satisfies certain regularity conditions, then the tree-structured filter bank can be shown to compute a wavelet transform. Such a tree-structured filter bank is called an octave-band filter bank. The analysis side of the octave-band filter bank calculates the forward wavelet transform and the synthesis side calculates the inverse wavelet transform. At this point, our motivation for studying UMD filter banks becomes apparent. Under the conditions stated above, a UMD filter bank can be directly linked to a wavelet decomposition. Thus, UMD filter banks can be used to both design and implement wavelet transforms. Chapter 2. Preliminaries 17 H0(z) |JVf TM voinj ui[n] y\[n] vi H | M TM ; • Mf - i [«] • j M TM FM-I(Z) — * — wi[n] Figure 2.11: The UMD filter bank (revisited). 2.3.13 UMD Filter Bank Implementation In principle, the design of a PR UMD filter bank amounts to decomposing the identity matrix into two factors with desired properties. These two factors are simply the polyphase matrices of the filter bank. Once we have determined the analysis and synthesis polyphase matrices E(z) and R(z), respectively, we are ready to proceed to the implementation of the filter bank. Of course, the filter bank could be realized by directly implementing the filtering operations in each of the polyphase matrices, but it is often beneficial to break the filtering process into a number of smaller and simpler cascaded stages. Consider for a moment the analysis side of the filter bank. Instead of implementing E(z) directly, we further decompose E(z) as follows E(z) = En.l{z)-Ex(z)E0(z). Each of the {Ej(z)} can then be taken to represent a single filtering stage in the final implementation as depicted in Figure 2.12. Similarly, R(z) can be decomposed to produce R(z) = Rm-.i(z) — Ri(z)Ro(z). This corresponds to the cascade realization of the synthesis polyphase matrix shown in Figure 2.13. In the event that R(z)E(z) = I, we have R(z) = E~l (z). Thus, we could choose R(z) as R(z)^E^(z)E7i(z)-E;\(z). This factorization results in a certain symmetry between the analysis and synthesis sides of the filter bank which can often be advantageous. Assuming that we want to realize the polyphase matrices with a number of cascaded blocks, what type of polyphase matrix factorization might we use? As mentioned earlier, a wavelet transform can be constructed from one or more instances of a single UMD filter bank building block. Thus, in order to generate a wavelet transform that is both reversible and ITI, one need only construct a UMD filter bank building block with both of these properties. Given that we would like to construct reversible ITI transforms, are there particular factorizations that help to achieve this goal? This question will be addressed at length in the next chapter. 2.4 Image Coding The particular application of reversible ITI wavelet transforms considered in this thesis is image coding. In the sections that follow, topics relevant to this application are discussed briefly. We begin with a short introduction to image coding and compression. Then, transform-based image compression is discussed along with the motivation for its use. Performance measures for both lossy and lossless compression are also described. We discuss wavelet transforms in the context of image compression, and explain why such transforms are particularly effective for coding purposes. 2.4.1 Image Coding and Compression The discipline of image coding studies alternative representations of images, usually with some specific purpose in mind. The goal may be to find a representation with less redundancy so that fewer bits are required to encode the Chapter 2. Preliminaries 18 -•- >-Ed*) Figure 2.12: Block cascade realization of the analysis polyphase matrix. -»-' Rd*) Figure 2.13: Block cascade realization of the synthe-sis polyphase matrix. Original Image Forward Transform Transform Coefficient Quantizer and Encoder Encoded Image y Transform Coefficient Decoder and Dequantizer Reconstructed Image Inverse Transform Encoder Channel or Storage Decoder Figure 2.14: General structure of a transform-based image compression system. image, or to add redundancy in order to facilitate error detection/correction for information transmitted over noisy channels. The former type of coding is referred to as image compression. In this thesis, our interests focus primarily on image compression. There are two general approaches to image compression. In the first approach, the sample values of the original image are encoded directly to generate the coded bitstream. Such an approach is often referred to as a spatial-domain approach. In the second method, a transformation is first applied to the samples of the original image and then the transformed data are encoded to produce the compressed bitstream. A system employing this philosophy is said to be transform based. For lossy compression, transform-based coders are almost exclusively used as they tend to have much better performance for fixed complexity. On the other hand, in the case of lossless compression, spatial-domain coders have traditionally been employed in the past. In recent years, however, coders employing reversible ITI transforms have been growing in popularity for lossless coding. For a more detailed treatment of image compression and signal coding techniques beyond that which we present here, the reader is referred to [108] and [77]. 2.4.2 Transform-Based Image Compression Systems The general structure of a transform-based image compression system is shown in Figure 2.14. In this diagram, x represents the original image, y denotes the compressed image, and x represents the reconstructed image obtained from decompression. The goal, of course, is to design a system so that the coded signal y can be represented with fewer bits than the original signal x. As indicated by the diagram, the compressed bitstream may be stored or transmitted. In the case of storage, compression has the benefit of reducing disk or memory requirements, and in transmission scenarios, compression reduces the bandwidth (or time) required to send the data. Rather than attempt to code the sample values of the original image directly, a transform-based coder first applies a transform to the image and then codes the resulting transform coefficients instead. The transform is used in an attempt to obtain coefficients that are easier to code. Many signals of practical interest are characterized by a spectrum that decreases rapidly with increasing frequency. Images are a good example of a class of signals with this property. By employing transforms that decompose a signal Chapter 2. Preliminaries 19 into its various frequency components, one obtains many small or zero-valued coefficients which correspond to the high-frequency components of the original signal. Due to the large number of small coefficients, the transformed signal is often easier to code than the original signal itself. As a practical matter, in order for a transform to be useful, it must be effective for a reasonably large class of signals and efficient to compute. Consider now the compression process. First, the transform of the image is calculated. The particular transform employed is chosen so as to pack the energy from the original signal into a small number of coefficients that can be more efficiently coded. The next step in the compression process is to quantize and code the transform coefficients. Quantization is used to discard transform coefficient information that is deemed to be insignificant. In the case of lossless compression, no quantization is performed since all transform coefficient bits are equally important. Finally, the quantized coefficients are coded to produce the compressed bitstream. The coding process typically exploits a statistical model in order to code symbols with a higher probability of occurrence using fewer bits. In so doing, the size of the compressed bitstream is reduced. Assuming that the transform employed is truly invertible, the only potential for information loss is due to coefficient quantization, as the quantized coefficients are coded in a lossless manner. The decompression process simply mirrors the process used for compression. First, the compressed bitstream is decoded to obtain the quantized transform coefficients. Then, the inverse of the transform used during compression is employed to obtain the reconstructed image x. 2.4.3 Wavelet Transforms for Image Compression Wavelet transforms have proven extremely effective for transform-based image compression. Since many of the wavelet transform coefficients for a typical image tend to be very small or zero, these coefficients can be easily coded. Thus, wavelet transforms are a useful tool for image compression. The main advantage of wavelet transforms over other more traditional decomposition methods (like the DFT and DCT) is that the basis functions associated with a wavelet decomposition typically have both long and short support. The basis functions with long support are effective for representing slow variations in an image while the basis functions with short support can efficiently represent sharp transitions (i.e., edges). This makes wavelets ideal for representing signals having mostly low-frequency content mixed with a relatively small number of sharp transitions. With more traditional transforms techniques like the DFT and DCT, the basis functions have support over the entire image, making it difficult to represent both slow variations and edges efficiently. The wavelet transform decomposes a signal into frequency bands that are equally spaced on a logarithmic scale. The low-frequency bands have small bandwidths, while the high-frequency bands have large bandwidths. This loga-rithmic behavior of wavelet transforms can also be advantageous. Since human visual perception behaves logarithmi-cally in many respects, the use of wavelet decompositions can sometimes make it easier to exploit characteristics of the human visual system in order to obtain improved subjective lossy compression results. Although wavelet transforms with many different characteristics are possible, orthogonal transforms with sym-metric finitely-supported basis functions are ideally most desirable for image compression. Orthogonality is beneficial as it ensures that transform coefficients do not become unreasonably large and also because it easily facilitates the selection of the most important transform coefficients in the sense of minimizing mean squared error. Symmetric basis functions are desirable in order to avoid undesirable phase distortion as a result of compression. If phase is not preserved, edges and lines can become severely distorted, resulting in poor subjective image quality. Moreover, the symmetric extension method for handling finite-length signals can only be applied to transforms with symmetric basis functions. This is yet another incentive for using transforms with symmetric basis functions. Unfortunately, in the case of two-band wavelet transforms, orthogonality, symmetry, and finite support can only be achieved in the trivial case of the Haar and other Haar-like transforms. For this reason, we usually choose to sacrifice orthogonality, and use biorthogonal transforms instead. In practice, this concession does not pose any serious problems. 2.4.4 Compression Performance Measures Obviously, in order to evaluate the performance of image compression systems, we need a way to measure compres-sion. For this purpose, the compression ratio (CR) metric is often employed and is defined as original image size in bits compressed image size in bits Chapter 2. Preliminaries 20 In some cases, the normalized bit rate (NBR) is more convenient to use, and is defined as the reciprocal of compression ratio. That is, j ^ g ^ _ compressed image size in bits original image size in bits Sometimes compression is instead quantified by stating the bit rate (BR) achieved by compression in bpp (bits per pixel). The bit rate after compression and compression ratio are simply related as BR = (bits/pixel for original image)/CR. In most cases, the author prefers to use the normalized bit rate and compression ratio metrics, since in order for the bit rate metric to be meaningful one must also state the initial bit rate. The normalized bit rate and compression ratio metrics, however, are meaningful if stated in isolation. In the case of lossy compression, the reconstructed image is only an approximation to the original. The difference between the original and reconstructed signal is referred to as approximation error or distortion. Although many metrics exist for quantifying distortion, it is most commonly expressed in terms of mean-squared error (MSE) or peak-signal-to-noise ratio (PSNR). These quantities are defined, respectively, as M-XN-l i yi—i yv—i M S E = I ^ £ E(*M-*M)2, and (2.28) M " ;=o ;'=0 p s n r = 2 0 1 0 8 1 0 ( 2 - 2 9 ) where x \ i , j] is the original image with dimensions M x N having P bpp, and x[i, j] is the reconstructed image. Ev-idently, smaller MSE and larger PSNR values correspond to lower levels of distortion. Although these metrics are frequently employed, it is important to note that the MSE and PSNR metrics do not always correlate well with image quality as perceived by the human visual system. This is particularly true at high compression ratios (i.e., low bit rates). For this reason, one should ideally supplement any objective lossy compression performance measurements with subjective tests to ensure that the objective results are not misleading. Generally speaking, distortion varies with the amount of compression. In other words, distortion is implicitly a function of rate (i.e., compression ratio). For this reason, plots (or tables) of distortion versus rate are often used to analyze lossy compression performance. Obviously, for any given rate, the lowest possible distortion is desired. In the case of lossless compression, the reconstructed image is an exact replica of the original image. In other words, the distortion is always zero. Since the distortion is zero, we only need to consider the rate achieved when analyzing lossless compression performance. Obviously, the lower the rate, the better is the compression performance. Despair is the price one pays for setting oneself an impossible aim. It is, one is told, the unforgivable sin, but it is a sin the corrupt or evil man never practices. He always has hope. He never reaches the freezing-point of knowing absolute failure. Only the man of goodwill carries always in his heart this capacity for damnation. —Graham Greene, The Heart of the Matter, 1948 21 Chapter 3 F r a m e w o r k s fo r Revers ib le I T I Wave le t /B lock T rans f o rms Wednesday night I fell asleep naturally, exhausted, but woke up with the most intense feeling of sadness ... fragments of dreams spinning into lonely wakefulness ... Do you know what it's like to be sad like that? As if your best friend had just died, except the sorrow is worse because that's not what happened, because even while people are leaving you you've always been alone. —Evelyn Lau, Runaway: Diary of a Street Kid, 1989 Overview The generalized reversible I T I transform (GRITIT) framework, a single unifiedframework f o r reversible I T I wavelet/block transforms, is proposed. This new framework is then used to study several previously proposed frameworks and their interrelationships. For example, the framework based on the overlapping rounding transform is shown to be a spe-cial case of the lifting framework with only trivial extensions. The applicability of the GRITIT framework f o r block transforms is also demonstrated. The generalized S transform (GST), a family of reversible I T I block transforms, is proposed. This family of transforms is then studied in detail. Throughout a l l of this work, particularly close attention is paid to rounding operators and their characteristics. 3.1 Introduction Although reversible ITI wavelet transforms can be constructed using ad-hoc methods, a more structured approach to the design of such transforms is clearly advantageous. By using a highly structured approach to transform design, we can produce more sophisticated transforms than is possible with ad-hoc methods. Thus, general frameworks for the construction of reversible ITI wavelet transforms are of great interest. In this chapter, we consider frameworks for the construction of reversible ITI wavelet transforms. We begin, in Section 3.2, by studying one of the key elements in such frameworks, namely rounding operators. This study of rounding operators will prove quite useful at various junctures in this thesis. In Section 3.3, we examine several pre-viously proposed frameworks. Then, in Section 3.4, we integrate and extend the key concepts from these frameworks in order to form a single unified framework, known as the generalized reversible ITI transform (GRITIT) framework. The GRITIT framework is then used, in Section 3.5, to study the relationships between the various other frameworks introduced in Section 3.3. Next, Section 3.7 demonstrates the applicability of the GRITIT framework to the design of reversible ITI block transforms. In this context, we propose and study the generalized S transform (GST), a new family of reversible ITI block transforms. Finally, in Section 3.8, we summarize the results of this chapter. In passing, we would like to note that some of our results presented in this chapter have also been published in [10,17,15,18]. Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 22 Figure 3.1: Identities for an integer-bias invariant rounding operator (x,y G Z, a € R, and Q is a rounding operator). 3.2 Rounding Operators As will soon become evident, rounding operators play an important role in reversible ITI wavelet transforms. There-fore, it is beneficial to devote some time to studying such operators. In what follows, we examine various matters concerning rounding operators. For example, we consider their basic properties, rounding error characteristics, and implementation. At times, we focus our attention on several common rounding operators (i.e., the floor, biased floor, ceiling, biased ceiling, truncation, biased truncation, and RAFZ functions). 3.2.1 Integer-Bias Invariance and Oddness Two properties that rounding operators may possess are integer-bias invariance and oddness, as defined by (2.13) and (2.14), respectively. First, let us consider integer-bias invariance. If a rounding operator has this property, the order of rounding and integer increment/decrement operations can be interchanged without affecting system behavior. In other words, we have the identities shown in Figure 3.1. Next, let us consider oddness. As noted earlier, if a round-ing operator possesses this property, it will preserve both symmetry and antisymmetry in signals. For example, if an antisymmetric signal is rounded with an odd operator, the resulting signal will also be antisymmetric. In some appli-cations, such as symmetric extension (to be discussed later), rounding operators that preserve both signal symmetry and antisymmetry are often desirable. For the reasons outlined above, the integer-bias invariance and oddness properties are both highly desirable. Un-fortunately, as it turns out, these properties are also mutually exclusive, as demonstrated by the proposition and ac-companying proof below. Proposition 1 (Mutual exclusivity of integer-bias invariance and oddness). A rounding operator cannot be both integer-bias invariant and odd. Proof. Let Q denote a rounding operator. Assume that Q is both integer-bias invariant and odd. Now, consider the quantity Q{k + \) where £ 6 Z. Using trivial algebraic manipulation and the integer-bias invariance property, we have Q ( k + i ) = Q ( 2 k + 1 - k - \ ) = 2 k + l + Q ( - k - \ ) . (3.1) From the oddness property, we can write Q ( k + \ ) = - Q ( - \ k + \ \ ) = - Q { - k - \ ) . (3.2) Combining (3.1) and (3.2), we obtain 2 k + \ + Q { - k - \ ) = - Q { - k - \ ) => 2 Q { - k - \ ) = - 2 k - l => Q { - k - \ ) = - k - \ . Thus, we have that Q ( - k - \) Z. Since, by definition, Q must always yield an integer result, the integer-bias invariance and oddness properties cannot be simultaneously satisfied by Q. • Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 23 3.2.2 Relationships Involving Rounding Functions In this section, we introduce some important relationships involving specific rounding operators (namely, the floor, biased floor, ceiling, and biased ceiling functions). To begin, we remind the reader of the very basic identities involving the floor and ceiling functions given by (2.1), (2.2), and (2.3). These identities are extremely useful and will be used repeatedly throughout this thesis. In what follows, we introduce some additional relationships involving rounding operators. First, we examine the relationship between the floor and ceiling of a quotient of integers. In this case, a simple identity can be shown to hold, as evidenced by the proposition and accompanying proof below. Proposition 2 (Relationship between floor and ceiling of quotient). For all x f Z and a l l y e Z except y = 0, the following identity holds: rVyl = L(*+y-i)/yJ- (3-3) Proof. One can easily confirm that, for y ^ 0, the mod function satisfies the relationships mod(x,y) = JC for 0 < x < y— 1, and mod(jt + y,y) = mod(jc,y). From the two preceding properties, we can deduce mod(— x,y) =y— 1 — mod(;c— l,y) fory^O. (3.4) Now we consider the expression on the right-hand side of equation (3.3). Using (2.10), we can write L(*+y-i)/yJ = L(*-i)/yJ + i <3-5) = ( x + y - 1 -mod(x- l,y))/y. Using (2.11) and (3.4), we can further manipulate (3.5) as follows: [ { x + y - l ) / y \ = (x + m o d { - x , y ) ) / y = rv>i-Thus, the identity given in (3.3) holds. • The identity given by the above proposition is practically useful, as it allows us to implement the ceiling of a quotient in terms of the floor of a quotient and vice versa. Next, we consider the relationship between the biased floor and biased ceiling functions. The relationship between these functions is, in fact, quite simple, and is given by the following proposition and accompanying proof. Proposition 3 (Relationship between biased floor and biased ceiling functions). The biasedfloor and biased ceiling functions are related as follows: {bfloor(a) — 1 i f a is an odd integer multiple of i (3.6) bfloora otherwise for all aeR. Proof. First, let us consider the case that a is an odd integer multiple of \ . In this case, a+ \ and a - 5 are both integer quantities. Thus, we have bfloora= [a + = a+ \ and bceila = [ a - 5 ] = a - \ . Combining these two equations, we obtain bceila = bfloor(a) - 1. Thus, we have shown that (3.6) holds when a is an odd integer multiple of \ . Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 24 Now, let us consider the case that a is not an odd integer multiple of | . In this case, we have bfloora= |a+2J = OL+\+AO and (3.7) bceila= [ a - i ] = a - i + A i , (3.8) where Ao and Ai are real and satisfy Ao € (—1,0) and Ai e (0,1). The intervals for Ao and Ai follow immediately from the definition of the floor and ceiling functions and the fact that the rounding error cannot be zero since a +1 and a - j cannot be integer quantities. Using simple interval arithmetic, we can deduce from (3.7) and (3.8) that the quantities bfloora and bceila must both be contained on the interval (a — j , a + \ ) . Since this interval has a width of one and is open at both ends, however, it cannot contain more than one integer value. Therefore, bfloora and bceila, which are integer quantities, must be equal. Thus, we have shown that (3.6) holds when a is not an odd integer multiple of j, completing the proof. • The identity given by the above proposition is useful in that it allows us to convert expressions using the biased floor function to expressions employing the biased ceiling function and vice versa. In practice, it is frequently necessary to round a real number of the form ^ where x £ Z, F e Z, and F > 1 (i.e., a dyadic rational number). That is, we often need to compute an expression of the form Q ( ^ r ) where Q is a rounding operator. For this reason, we desire some efficient means of evaluating such an expression for various choices of Q. Many of today's computing platforms utilize two's complement integer arithmetic. On such platforms, a very efficient method exists for computing the floor of a dyadic rational number. This method involves the arithmetic shift right (ASR) operation, and is given by the proposition and accompanying proof below. Proposition 4 (Relationship between arithmetic right shift and floor of quotient). Suppose that x is an integer represented in two's complement form with a word size o f P bits. Let asr(jc,n) denote the quantity resulting from the arithmetic right shift ofx by n bits. We assert that asr(x, n) = [jn J where 0 < n < P - \ . (3.9) Proof. Let a o , a \ , . . . ,a/>-i denote the P bits of the word representing x , where ao and a/>_i are the least- and most-significant bits, respectively. In two's complement form, the integer x is represented as x = - a P - l 2 p - l + - z £ 0 2 a i 2 i . Through the properties of modular arithmetic, one can use the preceding equation to show that mod(x,2") = X ? = 0 ' f o r 0 < « < P - 1. (3.10) From the definition of the arithmetic shift right operation, we can write (for 0 < n < P — 1) asr(x,n) = -ap^2P'X +aP.l (ZU-i-n21) + Sf=o 2 _" «<+«2'' = -aP.x ( 2 p - 1 - 2 ' " 1 + 2 p- '-") + X f - o 2 - % + „ 2 ' -= - « P _ 1 2 / ' - 1 - n - M f = 0 2 - " a / + „ 2 i . Using simple algebraic manipulation, (2.10), and (3.10), we can further rearrange the preceding equation as follows: asr(;c, n) = - a p ^ - 1 - " + S f " . 2 ; " a,+„2< - ^ a i + n 2 > = j n ( - a P - i 2 p - ' + S £ o fl«2') - £ m o d ( j c , 2 » ) _ -^mod(x,2") — 2 " = L#J-Thus, the identity given in (3.9) holds. (As an aside, we note that (3.9) also trivially holds for n > P. This case, however, is of relatively little practical interest.) • Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 25 Table 3.1: Error characteristics for various rounding operators (for F > 1) Operator Error Interval PAE MAE Exact For Large F Exact For Large F Exact For Large F floor bfloor ceiling l - ^ r M ] I 2F '2J [0,3^] r 1 2 F- 1-li 1 2' 2F J r 2F-1 2 f-li 1 2F ' 2F J l ~ 2' 2I r 2 f-l 2 f -l i I 7/' i 2 f ' -1 (-1,0] (_2> 2! [0,1) 2'-l 2'1 1 2 2 f-l 1 1 2 1 2'-l 2A'+1 1 4 2 f-l 2f+l \ 1 4 1 2 bceil trunc btrunc rafz [ 2>2> (-1,1) f_I i l I 2' 2l (-1,1) 1 2 2 f-l ~T I T 1 2 1 1 2 1 1 4 2F-\ 2F+\ Xx 1 4 1 2 1 4 1 2 Let us continue to assume a two's complement representation for integers. This being the case, we can use the results of Propositions 2 and 4 to derive the following expressions: LjrJ = asr(x,F), bfloor( ) = asr(* + H , F ) , T^l =asr(x + Af,F), bceil(^) = asr(jc-r-fl,F), Jasr(jc,F) forx>0 jasr(x+ M,F) forx<0, asr(jc + /Y,F) forjc>0 asr(x + B,F) for*<0, (asr(;c+ M,F) forx>0 I asr(x, F) for x < 0, and (3.11a) (3.11b) (3.11c) (3. lid) (3. lie) (3.1 If) (3-llg) where M = 2 F - \ , H - 2 F ~ X , B = 2 F _ 1 - 1, and asr(jc,n) denotes the arithmetic right shift of x by n bits. The above expressions are useful as they provide an efficient means for applying rounding operations to dyadic rational numbers. 3.2.3 Rounding Error Suppose that we must round real numbers of the form ^ where x € Z, F € Z, and F > 1 (i.e., dyadic rational numbers). Such numbers are associated with a fixed-point representation having F fraction bits (i.e., F bits to the right of the binary point). If we assume that, for both negative and nonnegative quantities, all 2 F representable fractional values are equiprobable, one can show that the error interval, peak absolute error (PAE), and mean absolute error (MAE) for each of the rounding operators under consideration are as listed in Table 3.1. The derivation of these expressions is given by the proposition below. Proposition 5 (Approximation error for various rounding operators). Consider the quantity where x is a random integer and F is a strictly positive integer constant. I f we suppose that frac ^ may assume all 2 F possible (distinct) values, then the error incurred by rounding the quantity using the the floor, biased floor, ceiling, biased ceiling, truncation, biased truncation, and RAFZ functions, is bounded as follows: - ^ < L # J - F < 0 , - ^ < b f l o o r ( ^ ) - £ < i , 0 < [ £ ] - £ < ^ , - \ < bceil ( £ ) _ £ < 2 ^ i , (3.12) - ^ r 1 < t runc(^) -^-< ^ i , _ i < btrunc (ifc) - < \, and Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 26 I f we further suppose that all 2 F possible values o/frac are equiprobable for ^ < 0 and are also equiprobable for > 0, then rounding jjr with the floor, biased floor, ceiling, biased ceiling, truncation, biased truncation, and RAFZ functions incurs the respective mean absolute errors: E(IUH-£l) = i ^ E( |bf loor(£)-£ | ) = J, E ( | [ # l - i H ) = ^ , E( |bce i l (£ ) -£ \ ) = \ , (3.13) E(| trunc(f)-^|) = ^ f , E(|btrunc(f) - £ | ) = \, and E(|rafe(^)-^|) = Proof. From (2.9) and (2.3), we can show that the rounding error incurred by the floor, biased floor, ceiling, and biased ceiling functions is, respectively, given by: L # J - # = ^ A (3-14) bf loor(^) -^ = 2 f"'-m o djr' +^, (3.15) r ^ - | _ ^ = ™ ^ i > and (3.16) bceil = m ° d ( 2 f " ' - f F ) - 2 F " ' - (3-H) Using the definitions of the truncation, biased truncation, and RAFZ functions given by (2.6), (2.7), and (2.8), respec-tively, and the identities (3.14)—(3.17), we can easily deduce: ^w)-W-{^n f Q T X < 0 t (3.18) » ™ * * ) - * - { ^ ^ f Q I X - < Q j (3.19) f , x , * fOTX>0 r a f z ^ - ^ H - ^ {oTx<Q <3-20> 2' Since mod(±;t,2F) and mod(2/7_1 ± x , 2 F ) can assume any integer value on [ 0 , 2 F - 1], we can easily deduce all of the error bounds in (3.12) from (3.14)-(3.20). For example, consider the case of the floor function. The rounding error in this case is given by (3.14). Clearly, the minimum value of the rounding error is — ^ p — , occurring when mod(;c,2F) = 2 F - 1. Similarly, the maximum value of the rounding error is 0, occurring when mod(*,2F) = 0. Thus, we have that - < [^.J - ^ < 0. The remaining error bounds in (3.12) can be shown to hold in a similar manner. Now, let us consider the mean absolute error expressions for the various rounding operators. Since, by assumption, all 2 F possible values for frac are equiprobable, we know that m o d ( ± x , 2 F ) and mod(2 / r-1 ± x , 2 F ) are both uni-formly distributed over the integer values on [0,2F -1]. Consequently, in the case of the floor, biased floor, ceiling, and biased ceiling functions, the mean absolute error can be found by computing the expected value over the 2 F possible absolute error values (associated with the 2 F distinct fractional parts) using (3.14)—(3.17). Straightforward analysis and algebraic manipulation yield the following: E(|L#J - w\) = E ( | = ^ | ) = jft 1^1 = S i r . (3-21) v w /=o E ( | b f l oo r (£ ) -£ | )=E( 2f-'-mod(2F-2 f - l 'I i=0 2 f - ' - i i 4' (3.22) Chapter 3. Frameworks for Reversible ITI Wavelet/Block Transforms 21 E ( | ^ l - # | ) = E ( | ^ ^ | ) = f 2 X |£| = £ T , and (3.23) i=0 2 f - l : ( | b c e i l ( ^ ) - ^ | ) ^ E ( | ^ - 1 ? 2 f ) - 2 f - 1 [ ) ^ ^ S i-2F~ l 4- (3.24) Let us now consider the mean absolute error for the truncation function. From (3.21), (3.23), and the definition of the truncation function in (2.6), it follows that E(| trunc(£) - jr\) = Pr(# < 0)E(| \ £ \ - $ \ \ § < 0) + P r ( £ > 0)E(| [^J - £ | | £ > 0). (3.25) Since, by assumption, all 2 F possible values of frac ^ are equiprobable for ^ < 0, we have from (3.23) that E ( | [ # l - ^ | l # < 0 ) = ^ f . (3.26) Similarly, since all 2 F possible values of frac ^ are assumed to be equiprobable for > 0, we have from (3.21) that • E ( | L ^ J - ^ | l ^ > 0 ) = Srf. (3-27) By substituting (3.26) and (3.27) into (3.25) and then observing that Pr(^- < 0) +Pr (^ > 0) = 1, we obtain E (|trunc(£) - £ | ) = P r ( £ < 0) (£=f) + P r ( £ > 0) (f^ f ) = £=f. Therefore, the identity for the truncation function in (3.13) holds. Consider now the mean absolute error for the biased truncation function. From (3.22), (3.24), and the definition of the biased truncation function in (2.7), it follows that E (|btrunc(£) - £ |) = P r ( £ < 0)E (|bceil(£) - £ | | jr < 0) (3.28) + P r ( £ > 0)E (|bfloor(£) - £ | | § > 0) . Since, by assumption, all 2 F possible values of frac ^ are equiprobable for ^ < 0, we have from (3.24) that E ( | b c e i l ( £ ) - £ | | £ < 0 ) . = £. (3.29) Similarly, since all 2 F possible values of frac are equiprobable for ^ > 0, we have from (3.22) that E ( | b f l o o r ( ^ ) - ^ | | ^ > 0 ) = i . (3.30) By substituting (3.29) and (3.30) into (3.28) and observing that Pr(-$r < 0) + P r ( ^ > 0) = 1, we obtain E(|btrunc(£) - £ | ) = i P r ( £ < 0) + ± P r ( ^ > 0) = \ . (3.31) Therefore, the identity for the biased truncation function in (3.13) holds. Using similar reasoning to that above, in the case of the RAFZ function, we can write: E ( | r a f z ( ^ ) - ^ | ) = P r ( ^ < 0 ) E ( | ^ J - ^ | | ^ < 0 ) + P r ( ^ > 0 ) E ( | ^ l - ^ | | ^ > 0 ) = 2 i&r) + 1 ( 2 ^ ) _ 2 f - l - -jT+r-Therefore, the identity for the RAFZ function in (3.13) holds. Thus, we have shown that all of the identities in (3.13) hold, completing the proof. • Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 28 (a) (b) Figure 3.2: The S transform. The networks for the (a) forward and (b) inverse transforms. 3.3 Previously Proposed Frameworks To date, numerous frameworks have been proposed for reversible ITI wavelet transforms. In the sections that follow, we introduce several of the more popular frameworks, namely the S+P transform, lifting, and overlapping rounding transform (ORT) frameworks. We begin, however, by introducing the S transform, a specific example of a reversible ITI wavelet transform. A brief examination of this transform is instructive as such provides additional insight into the evolution of frameworks for reversible ITI transforms. In recent years, these frameworks have evolved considerably, and consequently, it is now possible to design transforms that are much more sophisticated than the S transform. Notwithstanding its crudeness, however, the S transform is still useful in some applications. For this reason also, we elect to discuss the S transform herein. 3.3.1 S Transform One of the simplest and most ubiquitous reversible ITI transforms is the S (i.e., sequential) transform [28,108]. The S transform is essentially a reversible ITI approximation of the (linear) Haar wavelet transform [62]. According to [28], the S transform, as it is known today, appears to have originated from the work of Wendler [ 1.46] (which relates to the earlier work of Lux [96]). The S transform has become quite popular for lossless signal coding applications and has received considerable attention in the literature over the years (e.g., [110,28,1.18, .108,148,114]). Several slightly different definitions of the S transform exist in the literature. In what follows, we consider one of the more commonly used definitions (e.g., equations (1) and (2) in [114] and equation (3.3) in [ 148]). The S transform is associated with the UMD filter bank shown in Figure 3.2. The forward transform is computed by the analysis side of the filter bank (shown in Figure 3.2(a)), while the inverse transform is computed by the synthesis side of the filter bank (shown in Figure 3.2(b)). From Figure 3.2, we can see that the S transform can be characterized mathematically as follows. The forward transform splits the input signal x[n] into the lowpass and highpass subbands signals, yo[n] and yi [n], respectively, as given by yo[n] = {x[2n] + Jt[2n+ 1])J and (3.32a) y i [ n ] = x [ 2 n ] - x [ 2 n + l } . (3.32b) The inverse transform combines the lowpass and highpass subband signals, yo[n] and y \ [«], respectively, to yield the output signal x[n], as specified by *[2n] = yo[«] + Ls^ i Wj a n d ( 3 3 3 a) x [ 2 n + l ] = x[2n] - y i [n]. (3.33b) Although equations (3.32) and (3.33) assume signals to be of infinite length (i.e., boundary conditions for filtering are not considered), finite-length signals can be accommodated by using various extension strategies (e.g., periodic extension or symmetric extension). The treatment of finite-length signals will be discussed at length later in this thesis. Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 29 Figure 3.3: The S+P family of transforms. The networks for the (a) forward and (b) inverse transforms, where Q denotes the biased floor operation. The development of the S transform is based on two key observations. First, any two numbers can be unambigu-ously determined from their sum and difference. Second, the sum and difference of any two integers always have the same parity (i.e., the sum and difference are either both even or both odd). The first observation initially leads to the formulation of a transform that simply computes the pairwise sum and difference of samples. Although such a transform is reversible and ITI, we can further refine its definition by making use of the second observation above. Because the sum and difference of two integers must have the same parity, the least significant bit of either the sum or difference is redundant and may be discarded without losing information. Thus, we can choose to divide the sum by two and discard the fractional part of the result without any information loss. This is, in fact, exactly what the S transform does. Moreover, this explains how it is possible for the S transform to be reversible. A formal proof of the reversibility of the S transform is omitted here, but one can be found, for example, in [!]. For reasons yet to be discussed, one can trivially deduce the reversibility of the S transform directly from its associated filter bank shown in Figure 3.2. 3.3.2 S+P Transform Framework One of the first proposed frameworks for reversible ITI wavelet transforms was introduced by Said and Pearlman [114], and is known as the S+P (i.e., sequential plus prediction) transform framework. This framework can be viewed as a direct extension of the S transform. In fact, the S transform is an example of one specific transform that can be constructed using this framework. The S+P transform framework is associated with a filter bank of the form shown in Figure 3.3. The forward transform is computed by the analysis side of the filter bank (shown in Figure 3.3(a)), while the inverse transform is computed by the synthesis side of the filter bank (shown in Figure 3.3(b)). By comparing Figures 3.3 and 3.2, we can see that the transforms in the S+P family and the S transform share a similar structure. Mathematically, we can describe the S+P transform as follows. The forward transform splits the input signal x[n] Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 30 Table 3.2: Sets of predictor coefficients for the S+P transform framework Predictor OC-i Oo P - l A B l I 1 I 8 0 0 0 2 8 C Tfi 4 Ifi 1 16 6 Ifi into the lowpass and highpass subband signals, yo[«] and yi [n], respectively, as given by yo[n] = [| (x[2n] + x [ 2 n + 1])J and (3.34a) yi[n] = v 0 [ n } - [ t [ n } + {\, (3.34b) where V 0 [ H ] = x[2n] - x [ 2 n + 1], Li -1 t[n] = £ O . i { y o [ n - i - l ] - y o [ n - i ] ) - ^ W " " ' ] . i=io i=K and Lo, Z-i, and K are integers satisfying Lo < L\ and AT < — 1. The inverse transform uses the lowpass and highpass subband signals, yo[n] and yi [n], respectively, to form the output signal x[n] as follows: x[2n\ = yo[n) + |± (v0[n] + 1)J and (3.35a) x[2n +1] = x[2n] - v0[n], (3.35b) where M"]=yi[n]+[*["] +i\ and t[n] is as given above. From the above equations, we can see that the S+P transform framework is obtained by simply adding an extra processing step to the S transform. That is, this framework is generated by adjusting the highpass subband signal of the forward S transform with an extra prediction operation. The predictor is associated with the coefficients {a,} and {B,}. Three sets of predictor coefficients have been suggested by Said and Pearlman [114], as listed in Table 3.2. Of these, predictor A has the lowest computational complexity, and corresponds to the well known two-six (TS) trans-form [148]. In the degenerate case where all of the predictor coefficients are zero, the S transform is obtained. It is important to note that the synthesis side of the filter bank is associated with recursive filtering structures if any of the {P,} are nonzero. Although equations (3.34) and (3.35) assume signals to be of infinite length (i.e., boundary conditions for filtering are not considered), finite-length signals can also be accommodated. In order to handle finite-length signals, techniques such as periodic extension and symmetric extension (to be discussed later) can be used in some cases. This matter becomes somewhat more complicated when recursive filtering structures are employed. Moreover, the use of recursive structures has other important implications. All of these matters will be addressed later in this thesis. 3.3.3 Lifting Framework One of the most popular frameworks for reversible ITI wavelet transforms is based on the lifting scheme, a polyphase realization strategy for UMD filter banks. The lifting scheme, proposed by Sweldens [129], was originally presented as an alternative method for constructing biorthogonal ID two-band wavelet transforms [130, 131]. This scheme, however, has proven to be extremely useful in numerous other contexts as well. In particular, the lifting scheme has been shown to provide a means for constructing reversible ITI wavelet transforms. Essentially, the lifting scheme is a polyphase realization strategy for UMD filter banks that employs ladder net-works for polyphase filtering. The lifting realization of a ID two-channel UMD filter bank has the general form shown Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 31 x[n] FPT Ao(z) .yi["i (a) yow ^2X-2(z) <5-Ao(z) • . A-. IPT x[n] (b) Figure 3.4: The lifting realization of a linear ID two-band wavelet transform. The network for the (a) forward and (b) inverse transforms. ykW ^ xkm (+) A(z) xi-k["] yi-k[n] yi-k[n] *\-k[n] (a) (b) Figure 3.5: Two ladder steps, (a) A ladder step and (b) its inverse. in Figure 3.4 (where FPT and IPT denote forward and inverse polyphase transforms, respectively). From the diagram, we can see that the polyphase filtering is performed by a ladder network and some additional amplifiers. The filters {A*} are referred to as lifting-step filters, and each such filter constitutes what is called a lifting step. The amplifiers, associated with the gains {st}, each constitute what is called a scaling step. The lifting realization is a particularly interesting one, primarily because of its use of ladder networks. Such networks have the remarkable property that they can be made to maintain their invertibility even in the presence of certain types of quantization error, particularly the rounding error introduced by finite-precision arithmetic. To see why this is so, consider the two ladder steps shown in Figure 3.5. Clearly, if exact arithmetic is employed, these two networks invert one another. Suppose now that the filtering operations (associated with A(z)) are implemented using finite-precision arithmetic and some roundoff error is incurred. When we cascade these two networks together, both ladder-step filters are presented with the same input. Since the two filters are identical (and assumed to use the same implementation strategy), they will both incur the same rounding error, and their outputs will be identical. Therefore, whatever value is added by the adder in the first network will be subtracted by the adder in the second network. Consequently, the two networks continue to invert one another, even in the presence of rounding error. It is not difficult to see that we can apply this argument repeatedly in order to show that a ladder network with any number of ladder steps can maintain its invertibility even in the presence of rounding error. In other words, such networks are fundamentally reversible in nature. Although the lifting realization has become quite popular recently, the idea of reversible ladder-based polyphase networks is not new. Such networks were first proposed by Bruekers and van den Enden [36], and later studied in some detail by others [81,82, 137]. In spite of the interest in the lifting realization, its application to reversible ITI wavelet transforms was not seen until later. In their seminal work, Calderbank et al. [40] first proposed using the lifting realization as a framework for con-structing reversible ITI ID two-band wavelet transforms. In order to create a reversible ITI version of a linear wavelet Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 32 A(z) (a) AW *[»] yi-t[n (b) Figure 3.6: Modifying a lifting step to map integers to integers, (a) Before modification, (b) After modification. .yo[n] FPT yim. Qo 4> Qix-Aix-dz) i Gi Mz) i *a -2 (z ) (a) (b) I 1 Mz) l Qo so .yiw IPT Figure 3.7: The lifting realization of a reversible ITI ID two-band wavelet transform. The networks for the (a) forward and (b) inverse transforms. transform, we first find its lifting realization, which has the form shown in Figure 3.4. Then, we eliminate any scaling steps in the forward transform having a non-integer gain factor, by either discarding these steps or converting them to equivalent lifting steps if possible. Finally, we transform each lifting step in the resulting system by introducing a rounding operation Q at the output of its corresponding filter, as shown in Figure 3.6. The above process yields a system of the form depicted in Figure 3.7. Clearly, the polyphase filtering is performed using lifting (i.e., L-type) and scaling (i.e., S-type) networks of the form shown in Figure 3.8. Since each lifting and scaling step in the forward transform is ITI in nature, the resulting transform is ITI. Moreover, the transform is reversible, since the lifting and scaling steps are also reversible. Any (reasonable) rounding operator may be used for Q. Techniques similar to the one described above have also been suggested by other researchers. For example, similar ideas for constructing reversible ITI wavelet transforms have been proposed by Chao et al. [44] and Dewitte and Cornells [50]. The lifting framework extends easily to the case of M-band wavelet transforms (e.g., as described in [I]). 3.3.4 Overlapping Rounding Transform (ORT) Framework Although the lifting framework has become quite popular since its introduction, other frameworks for reversible ITI wavelet/subband transforms have also been suggested. One such framework is the overlapping rounding transform Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 33 Xj[n] ym (a) (b) Figure 3.8: Networks used for polyphase filtering in the lifting framework. Networks for (a) L-type and (b) S-type operations. -> «-FPT To Ti TN-I (a) • — 'N-I r - i IPT *[»] (b) Figure 3.9: The ORT realization of a reversible ITI two-band wavelet/subband transform. The networks for the (a) forward and (b) inverse transforms. (ORT) framework, as proposed by Jung and Prost [79,78]. The ORT framework is a polyphase realization strategy for UMD filter banks that employs networks with a par-ticular structure for polyphase filtering. The ORT realization of a ID two-channel UMD filter bank has the general form shown in Figure 3.9. The forward transform is computed by the analysis side of the filter bank (shown in Fig-ure 3.9(a)), while the inverse transform is calculated by the synthesis side of the filter bank (shown in Figure 3.9(b)). As the diagram illustrates, polyphase filtering is performed by a cascade of two-input two-output reversible ITI net-works (associated with the {T}^1). Only eight types of reversible ITI networks can be employed, with the forms shown in Figure 3.10. The forward transform can utilize any number of A-, B-, C-, D-, and E-type operations in any order. The inverse transform is then formed by using the inverse networks corresponding to those used in the forward transform. The necessary networks for the inverse transform are obtained by making the following observations. The inverse of the A-, B-, and C-type networks are, respectively, the A - 1 - , B - 1 - , and C _ 1-type networks, as depicted in Fig-ure 3.10. The inverse of a D-type network is another D-type network with the opposite shift (i.e., in the diagram, K is replaced by —K in order to obtain the inverse structure). An E-type network is self inverting. Consider, for a moment, a transform constructed with the above framework. Clearly, any such transform is re-versible, since the FPT and each of the networks used for analysis polyphase filtering are reversible. Furthermore, since the FPT and each of these networks are ITI, the overall transform is also ITI. Although originally proposed in the context of two-band wavelet/subband transforms, the ORT framework can also be easily extended to the M-band case (e.g., as described in [80]). 3.4 Generalized Reversible ITI Transform (GRITIT) Framework In the preceding sections, we considered several previously proposed frameworks for reversible ITI wavelet transforms (namely, the S+P transform, lifting, and ORT frameworks). These frameworks utilize a number of key ideas, but only a subset of these ideas is employed by each individual framework. Clearly, it would be beneficial to have a single unified framework that exploited all of these ideas at once. Having such a tool at our disposal, we could more easily study different frameworks and the relationships between them. To this end, we have integrated the key ideas from existing frameworks and extended them somewhat in order to create a single unified framework for reversible ITI transforms known as the generalized reversible ITI transform (GRITIT) framework. In our work, we explicitly consider the most Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 34 xi[n xi[n] H(z) l-H(z) •<t> L-J yo[n] (a) L-J yi[r>] yoM H(z) ± 1 0 - L-J yiM (c) yo[«] (e) Ar,-[n] (g) yoM yi [« ] yoM yoM y i [ " ] *o[n] n xi[n\ (b) l±tf (z) *o[n] ±1 0 > (d) <tHn *i[n] (f) JCOH yoM Jti[n (h) Figure 3.10: Networks used for polyphase filtering in the ORT framework. Networks for (a) A-type, (b) A '-type, (c) B-type, (d) B~'-type, (e) C-type, (f) C_ 1-type, (g) D-type, and (h) E-type operations. Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 35 general case of D-dimensional M-band wavelet transforms (where D > 1 and M > 2). As we shall see later, the GRITIT framework can also be used as a framework for reversible ITI block transforms. With the GRITIT framework, a transform is constructed using a number of primitive reversible ITI operations. Given the operations employed in the forward transform, one can easily construct the inverse transform through the stepwise inversion of each operation in the forward transform. That is, the inverse transform is formed by applying, in reverse order, the inverse of each operation in the forward transform. In what follows, we will introduce the primitive reversible ITI operations employed by the GRITIT framework, and study this framework in more detail. As we shall see, in the case of wavelet transforms, this framework is fundamentally based on the polyphase realization of a UMD filter bank. 3.4.1 Primitive Reversible ITI Operations As suggested above, a number of primitive reversible ITI operations are employed by the GRITIT framework. Using these building blocks, we can construct both wavelet and block transforms. Six distinct types of operations are em-ployed, known as the split, join, displace, exchange, shift, and scale operations. Below, we will describe each of these operations in more detail. 3.4.1.1 Split Operation The split operation is associated with the 1-input M-output network shown in Figure 3.11(a). This operation decom-poses the input signal x[n] into M polyphase components {v^n]}^ 1 , where the polyphase decomposition is per-formed with respect to the the sampling matrix M and corresponding coset vectors {wi,-}^ 1 , and M = |detM|. In other words, the split operation simply computes a FPT. As a matter of notation, we denote an operation of this type as T ( M , [ m o mi • •• m M _ i ]). xhe split operation is linear and shift variant. Since the downsampling and shift operations are ITI, the split operation is also ITI in nature. The inverse of a split operation is a join operation (to be denned next). (More specifically, in terms of notation yet to be defined, ¥ ~ x ( M , [™*o « M - ™ » M - I ] ) = J ( M , [mo « M ••• m ^ - i ])). 3.4.1.2 Join Operation The join operation is associated with the M-input 1-output network shown in Figure 3.11(b). This operation sim-ply synthesizes a signal y[n] from its M polyphase components {^/[n]}^1. In other words, the join operation simply computes an IPT. As a matter of notation, we denote an operation of this type as 3 ( M , [ m o "M ••• « * M - I ] ) . The join operation is linear. Since the upsampling, shift, and addition operations are ITI, the join operation is also ITI in nature. The inverse of a join operation is a split operation. More specifically, J ~ X ( M , \ m o m i - ™-M-\)) = T{M,[MO ™M - ™»M-l]). 3.4.1.3 Displace Operation The displace operation is associated with the M-input M-output network depicted in Figure 3.11(c). As a matter of notation, the network inputs and outputs are denoted as {*,-[n]} and {v,[n]}, respectively. In order to simplify the diagram, only the Kth output is shown. Al l of the other outputs are directly connected to their corresponding inputs (i.e., y,[n] = x;{n] except for i = K). The Kth output (i.e., yn[«]) is generated by adding an adjustment value to the Kih input (i.e., *K[TI]). This adjustment value is calculated by summing filtered versions of one or more inputs (i.e., {^ •[nJj^ Q1) and possibly the ATth output (i.e., yjtfn]), and then rounding the result with the operator Q. If s is odd, the sign of the adjustment value is also inverted. Finally, the Kth output is formed by adding the adjustment value to the Kth input. Mathematically, we have y . r n ] = f*.-["] + fi(*[n]*3'i[n]+X^o1fl/[n]*JC/[n]) f o r i = K ' \ x i [ n ] otherwise. As a matter of notation, we denote an operation of this type as L(K,Q,s,[A0(z)Ai(z) - AM^(z)],B{z)). (3.36) In order to avoid delay-free loops (which are physically unrealizable), B(z) must have a constant term of zero. Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 36 The inverse of a displace operation is another displace operation. More specifically, L ~ l { K , Q , s , [Ao(z) - ,B(z)) = L ( K , Q , 1-s,\A'0(Z) A\{*) - A _^,(*)] ,B'(Z)) where B ' { z ) = A K { z ) and A-(z) = |A;(z) foriVAT I B ( z ) for i = fori = 0 , l , . . . , M - l . The displace operation is shift invariant, provided that Q is shift invariant. If Q is nonlinear (which is most fre-quently the case in practice), the displace operation is also nonlinear. One can easily confirm that, for any (reasonable) choice of Q, the displace operation is reversible. Since the rounding operator Q always yields an integer result (and the set of integers is closed on addition), the displace operation is ITI in nature. If all of the filter coefficients (i.e., the coefficients of the { A ^ z ) } ^ 1 a n d B ( z ) ) are integer valued, then no rounding operation is required in order to map integers to integers, and Q can simply be made the identity. In this case, we use the symbol 0 to indicate that no rounding operator is employed (e.g., as in £(1,0,1, [o l] ,0)). The displace operation can be viewed as a generalization of the lifting operation employed in the lifting framework. The primary difference between the displace and lifting operations is that the displace operation has a provision for feedback. That is, the displace operation is associated with a recursive filtering structure if B(z) ^ 0. (In effect, the lifting scheme always has B ( z ) = 0.) In the recursive case, however, care must be taken to ensure the existence of a causal inverse. For example, in the ID case, one could choose B(z) to have either strictly positive powers of z or strictly negative powers of z. If we neglect the effects of the rounding operator Q, the displace operation is linear, and therefore, its behavior can be characterized by a transfer matrix. T h i s M x M matrix only has nonzero entries on the main diagonal and in the Kth row. Each of the entries along the main diagonal is one, except (possibly) for the entry in the Kth row. The entries along the Kth row are such that the ( K , l ) t h entry has the form 1 + ^ g ^ . f ^ i f l = K and ^ I g ^ f otherwise. In other words, the matrix has a form resembling the following: 1 ( - l )Mo(« ) (-l)'itr-iW l+(- l )M g (z) (-QMjf+ifr) (-1)MM- ;(«) \-B(z) "• r=5(i) l-B(z) r^B^i) 1 3.4.1.4 Exchange Operation The exchange operation is associated with the M-input M-output network illustrated in Figure 3.11(d). As a matter of notation, the network inputs and outputs are, respectively, denoted as {•*/[«]} and {y,[n]}. In order to simplify the diagram, only the Kth and Lth outputs and inputs are shown. All of the other outputs are directly connected to their corresponding inputs (i.e., y,-[n] = */[n] except for i € { K , L } ) . The exchange operation simply swaps the signals in the Kth and Lth channels. Mathematically, we have { XL,{n\ for i = K x K [ n ] for i = L Xi[n] otherwise. As a matter of notation, this type of operation is denoted as £(AT, L). The exchange operation is linear and shift invariant. Moreover, one can see that this operation is trivially ITI in nature. Lastly, an exchange operation is self inverting. That is, £ - 1 { K , L ) = £ (^ ,L) . Since the exchange operation (i.e., £(#,L)) is linear, it can be characterized by a transfer matrix. This M x M matrix has all of its entries on the main diagonal equal to one, except the Afth and Lth entries which are zero. All of the off-diagonal entries are zero, except for the (AT,L)th and (L, AT)th entries which are one. Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 37 3.4.1.5 Shift Operation The shift operation is associated with the M-input M-output network depicted in Figure 3.11(e). As a matter of notation, the network inputs and outputs are, respectively, denoted as {*/[n]} and {y,[n]}. In order to simplify the diagram, only the Kth input and Kth output are shown. Each of the remaining outputs is directly connected to its corresponding input (i.e., y,[n] = jc,-[n] except for i = K). The shift operation forms the ATth output by translating the Kth input by the integer vector m . Mathematically, we have . _ i x i [ n — m] fox i = K ' \ x i [ n ] otherwise. As a matter of notation, we denote this type of operation as T ( K , m ) . This operation is linear and shift invariant. The inverse of a shift operation is another shift operation. More specifically, T~i ( K , m ) = T ( K , —m). Since the shift operation (i.e., T ( K , m ) ) is linear, it can be characterized by a transfer matrix. This M x M matrix is diagonal and has all of its diagonal entries equal to one, except for the Kth entry which is z ~ m . In other words, the matrix has a form resembling the following: '1 1 1 3.4.1.6 Scale Operation The scale operation is associated with the M-input M-output network illustrated in Figure 3.11(f). The network inputs and outputs are, respectively, denoted as {*/[n]} and {yi[n]}. TO simplify the diagram, only the Kth output is shown. Each of the remaining outputs is directly connected to its corresponding input (i.e., y,[n] = xt[n] except for i = K). Evidently, the scale operation forms the Kth output (i.e., y^n]) by simply applying a multiplicative (scalar) gain s to the Kth input (i.e., XK[n]). Mathematically, we have sxj[n] for i = K Xi[n] otherwise. As a matter of notation, we denote this type of operation as S ( K , s ) . The scale operation is linear and shift invariant. The inverse of a scale operation is another scale operation. More specifically, S ~ l { K , s ) = S { K , s ~ l ) . Clearly, a scale operation is only invertible if its associated gain s is nonzero. This type of operation is ITI in nature if the gain 5 is an integer. For this reason, any scale operations used in the computation of a forward transform must employ integer gain factors. Since the scale operation (i.e., S ( K , s ) ) is linear, it can be characterized by a transfer matrix. This M x M matrix is diagonal, and has all of its diagonal entries equal to one, except for the Kth entry which is s. In other words, the matrix has a form resembling the following: n i i s 1 1 Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 38 i M T , z m « - ' (a) yiM yu-l[n] xi [n] XM-\[n] T M T M T M 2~m\f j (b) J^f-t-l [n] %-l[n] A 0 ( z ) A , (2 ) Af - iW A J C + I ( * ) AM- 1(2) 7 t - i ) s B(z) xK[n] yK[n] (d) yn[n\ (e) yn[n] (f) (c) Figure 3.11: The six types of operations used in the GRITIT framework. The (a) split, (b) join, (c) displace, (d) ex-change, (e) shift, and (f) scale operations. Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 39 x[n] Split 7b Analysis Polyphase Operations (a) yo[n] yw-iH Synthesis Polyphase Operations (b) Join Figure 3.12: The general structure of a reversible ITI wavelet/subband transform using the GRITIT framework. The networks for the (a) forward and (b) inverse transforms. 3.4.2 Reversible ITI Wavelet Transforms Having introduced the primitive reversible ITI operations employed by the GRITIT framework, we are now in a posi-tion to explain how such operations can be utilized in order to construct reversible ITI wavelet transforms. As explained earlier, the basic building block for wavelet transforms is the UMD filter bank. Consequently, in order to synthesize a reversible ITI wavelet transform, we simply need to construct a reversible ITI UMD filter bank. Fortunately, this can be easily accomplished via the GRITIT framework. With the GRITIT framework, we can readily construct D-dimensional M-channel UMD filter banks based on polyphase techniques. Such filter banks are realized by employing the general structure shown in Figure 3.12. The analysis side of the filter bank, shown in Figure 3.12(a), is comprised of a split operation (i.e., fP) followed by one or more polyphase filtering operations (i.e., {7/}/!jol), a n ^ serves to decompose the input signal x[n] into M subband signals {y^n]}^ 1 . The polyphase filtering operations {Ti}^~01 are chosen as displace, exchange, shift, and scale operations (as desired). The synthesis side of the filter bank, shown in Figure 3.12(b), generates the output signal x[n] from its corresponding subband signals (i.e., {y^n]}^ 1)- This is accomplished via the stepwise inversion of each operation from the analysis side. (Recall that the inverse of a split operation is a join operation (i.e., fP - 1).) Since all of the operations employed (i.e., ( P , T ~ l , {7}}) are reversible, one can easily see that the overall transform is also reversible. Furthermore, since each operation on the analysis side maps integers to integers, the transform is also ITI. 3.4.3 Reversible ITI Block Transforms As suggested previously in Section 3.4, the GRITIT framework is applicable not only to wavelet/subband transforms but also block transforms as well. Block transforms are associated with M-input M-output networks. In the case Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 40 To Ti yo Xl y\ XM-l ; ; ; yM-(a) yo T-l 'o xo yi x\ y M - i XM-(b) Figure 3.13: The general structure of a reversible ITI block transform using the GRITIT framework. The networks for the (a) forward and (b) inverse transforms. of block transforms, each of the inputs and outputs is comprised of a single sample and not a sequence (as in the wavelet/subband transform case). For this reason, some minor modifications are required to adapt the GRITIT frame-work to the block transform case. First, we discard the split, join, and shift operations, as such operations are only meaningful for sequences. This leaves us with the displace, exchange, and scale operations. Next, we redefine these remaining operations to operate on individual sample values, as opposed to sequences. This redefinition is trivial for the exchange and scale operations. In the case of the displace operation, we restrict all of its associated filters to have constant transfer functions. This, in effect, replaces the filters by simple amplifiers. With the aforementioned modifications in place, we can employ the GRITIT framework to construct reversible ITI block transforms. This is accomplished by using the general structure shown in Figure 3.13. The networks in Figures 3.13(a) and 3.13(b), compute the forward and inverse transforms, respectively. The inputs and outputs of the forward transform are denoted, respectively, as { A T ; } ^ ' a n d ( ^ i / M ) 1 -By using the GRITIT framework, we can construct reversible ITI versions of linear block transforms such as the DCT and DFT. In fact, similar ideas have been recently proposed by a number of researchers [104,63,150]. Later, we will consider a particular family of block transforms in order to demonstrate the applicability of the GRITIT framework in the block transform case. 3.4.4 Practical Considerations Concerning Rounding As should now be apparent, the displace operation plays a crucial role in the construction of reversible ITI wavelet/block transforms using the GRITIT framework. This operation is associated with the computational structure shown ear-lier in Figure 3.11(c). The displace operation employs a rounding operator, denoted as Q in the diagram. Since any (reasonable) rounding operator can be selected for use in a displace operation, it is only natural to consider the impli-cations of such a choice. In what follows, we examine the impact that this choice has on computational complexity and rounding error for several common rounding operators, namely, the floor, biased floor, ceiling, biased ceiling, truncation, biased truncation, and RAFZ functions. Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 41 Impact of Rounding Operator on Computational Complexity As one might suspect, the choice of rounding operator can affect the computational complexity of a displace operation. Consider, again, the network associated with such an operation, as illustrated in Figure 3.11(c). This network simply computes / M-I \ yic[n] ^xK[n} + {-iyQU[n}*yK[n]+^ai[n}*xi[n]\ (3.37) where b[n] = Z r x B ( z ) and a,[n] = Z ~ x A t ( z ) for i = 0,1,... , M - 1. Without loss of generality1, we assume that all of the filter coefficients (i.e., the samples in the sequences {«/[«•] j / lo ' and b[n}) can be expressed exactly as dyadic rational numbers (i.e., numbers of the form where x 6 Z , F e Z , and F > 1). Provided that F is chosen large enough, we can always define Aj{z) = 2 F A j ( z ) forj = 0 , l , . . . , M - l and B{z)=2FB(z) such that and b[n] are integer sequences, where a,[n] = Z ~ x A [ ( z ) and b[n] = Z ~ x B ( z ) . Often, in practice, we need to compute the expression given by (3.37) using integer arithmetic. In such cases, we typically calculate yx[n] =xK[n] + ( - l ) 'Q ^ *)*[»] + £ &M *Xi[n] Thus, we must compute an expression of the form Q ( £ ) (3-38) where x € Z. In Section 3.2.2, we devised an efficient means for the computation of such an expression, for the cases of the floor, biased floor, ceiling, biased ceiling, truncation, biased truncation, and RAFZ functions. In particular, we derived the expressions given by (3.11). By examining these expressions, we can see that the floor function incurs the least computation (i.e., one shift). The biased floor, ceiling, and biased ceiling function require slightly more computation (i.e., one addition and one shift in total). Finally, the truncation, biased truncation, and RAFZ functions are the most complex due to their behavioral dependence on the sign of the operand. Impact of Rounding Operator on Rounding Error When constructing reversible ITI versions of linear transforms, one must remember that the effects of rounding error can be significant. The objective is to produce a (nonlinear) ITI mapping that closely approximates some linear transform with desirable properties. In order to preserve the desirable properties of the parent linear transform, it is usually beneficial to minimize the error introduced by rounding intermediate results to integers. In the GRITIT framework, rounding is performed exclusively in displace operations. The displace operation is associated with the network shown in Figure 3.11(c). If a (forward) transform only utilizes a single displace operation, the associated rounding error will be relatively small and its impact minimal. More typically, however, multiple dis-place operations are employed. In this case, rounding error may be more of a concern. Therefore, in some applications, it may be desirable to choose the rounding operator in order to minimize rounding error. Suppose that we must round numbers of the form where x 6 Z, F £ Z, and F > 1 (i.e., dyadic rational numbers), as in (3.38). In Section 3.2.3, we considered the rounding of such numbers, and characterized the error interval, MAE, and PAE for several common rounding operators (i.e., the floor, biased floor, ceiling, biased ceiling, truncation, biased truncation, and RAFZ functions). The results were summarized in Table 3.1. By examining the table, one can easily see that, for F > 2, the biased rounding operators have the smallest MAEs and PAEs. Consequently, the use of the biased operators has the advantage of reduced rounding error. It is important to note, however, that the preceding assertion does not hold if F = 1. (Typically, we would have F = 1 when the filter 1 Any given real number can be approximated with arbitrary accuracy by an appropriately chosen dyadic rational number. Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 42 uo[n\ A2(z) X (a) V2 TTj VN-i\n uoln\ A,(z) A2W (b) ± vi [n] v2[n) v f - l [ " Figure 3.14: Combining multiple displace operations that modify the same channel into a single displace operation, (a) Original network with multiple displace operations, (b) Network obtained after combining multiple displace oper-ations into a single displace operation. coefficients associated with a displace operation are all integer multiples of one half.) If F = 1, all of the rounding operators under consideration have the same PAE (i.e., \) and the same MAE (i.e., \) , and consequently, all of these operators are equally good in terms of rounding error. In practice, the case of F = 1 is not so unlikely to occur. Consequently, it is important to note the peculiar behavior in this case. In fact, some well known papers on reversible ITI transforms have overlooked this fact (and used a bias of one half when it is not beneficial to do so). Reducing Rounding Error By Combining Displace Operations When constructing reversible ITI wavelet/block transforms, we may occasionally encounter a network with several successive displace operations that all modify the same channel. In such a situation, in order to reduce rounding error, we can combine these displace operations into a single one. For example, suppose that we have a sequence of displace operations, all of which modify the Oth channel, as shown in Figure 3.14(a). Although we could naively perform rounding at the output of each filter, as shown in Figure 3.14(a), the rounding error can be significantly reduced by grouping the filters together and performing rounding only once, as shown in Figure 3.14(b). Clearly, in the absence of rounding, both networks are equivalent. Therefore, both networks approximate the same linear transform. In the first case, we have N - l displace operations, whereas in the second case, we have only one displace operation. By reducing the number of rounding operations, however, we can reduce the overall rounding error in the system. 3.4.5 GRITIT Realization of Wavelet/Block Transforms Earlier, we introduced the GRITIT framework and showed how the primitive reversible ITI operations associated with this framework can be used to obtain general structures for reversible ITI wavelet/block transforms. We have yet to discuss, however, how one might choose the specific operations and corresponding parameters for a particular transform. That is, so far, we have not considered the transform design problem. We will now turn our attention to this matter. The most common design technique is of an indirect nature. That is, we do not directly generate a reversible ITI transform. Instead, we first construct a linear wavelet/block transform with desirable properties. Then, we find a reversible ITI approximation of this transform. As we shall see, this design approach is essentially equivalent to a matrix factorization problem. The particulars of the factorization problem differ, however, in the wavelet and block transform cases. Wavelet Transforms In the wavelet transform case, we first choose a specific polyphase decomposition. This defines the split operation to be used. Next, we must determine the polyphase operations associated with the transform. To accomplish this, we simply need to decompose the analysis polyphase matrix into factors of the forms associated with the relevant Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 43 GRITIT operations (i.e., the displace, exchange, shift, and scale operations). That is, we must decompose the analysis polyphase matrix E(z) as E(z) = EN-l(z)..-El(z)E0(z) (3.39) where the {Et(z)} axe transfer matrices of the forms associated with the displace, exchange, shift, and scale opera-tions. (The forms of such transfer matrices are specified in Section 3.4.1.) Once we have determined the factorization in (3.39), the polyphase operations in the forward transform are obtained by simply selecting the operations corre-sponding to each of the matrix factors (i.e., the {22;}). The inverse transform is trivially formed by the stepwise inversion of each of the operations in the forward transform. Often, we are interested in wavelet transforms that are associated with FIR UMD filter banks. In this case, the corresponding analysis polyphase matrix E(z) has Laurent polynomial entries. Typically, in such a scenario, we want to obtain polyphase operations that are associated with FIR filters. That is, we want to obtain a factorization with Laurent polynomial matrix factors. In the ID case, in order to accomplish this, the factorization in (3.39) can be performed using a matrix Euclidean algorithm (e.g., as described in [49]). Provided that the transform is appropriately normalized, a solution to the factorization problem will always exist. This follows from the work of Daubechies and Sweldens [49] in which it was demonstrated that any (appropriately normalized) biorthogonal ID two-band wavelet transform associated with finitely-supported wavelet and scaling functions admits a lifting factorization with Laurent polynomial matrix factors. In the general D-dimensional case (where D > 1), this factorization problem is somewhat more complex, and a solution may not necessarily exist. For example, Kalker and Shah [SI] have noted that not all 2D UMD filter banks have a ladder realization. In some cases, however, a solution must exist as a consequence of a mathematical result known as Suslin's stability theorem [128]. To solve the factorization problem in the general D-dimensional case (assuming a solution exists), one can employ algorithms such as those proposed by Park and Woodburn [106] and Tolhuizen et al. [137]. Block Transforms In the block transform case, we simply need to decompose the forward transform matrix into factors having the forms associated with the relevant GRITIT operations (i.e., the displace, exchange, and scale operations). Let us denote the forward transform matrix as T. (We assume that T is a real matrix.) Then, we must decompose T as follows: T = TN-l---TlT0 (3.40) where the {I;} are matrices of the forms associated with the displace, exchange, and scale operations. Once we have determined the factorization in (3.40), the operations in the forward transform are obtained by simply selecting the operations corresponding to each of the matrix factors (i.e., the {Ti}). The inverse transform is trivially formed by the stepwise inversion of each of the operations in the forward transform. The above factorization process can be performed by simple Gaussian elimination. In order to avoid an unneces-sarily large number of factors, one might wish to employ a slightly more sophisticated technique like that proposed by Hao and Shi [63]. 3.4.6 Variations on the GRITIT Framework The GRITIT framework, as described previously, constitutes a very powerful tool for the study and construction of reversible ITI wavelet/block transforms. In the interest of simplicity, we chose to describe this framework in the most basic form suitable to our needs herein. We would be remiss, however, if we failed to note that many variations on this framework are possible. In this thesis, we are interested exclusively in reversible ITI transforms that approximate linear wavelet/block transforms. For this reason, the displace operation is based on linear filtering operations. One could, however, just as easily employ other types of (possibly nonlinear) operations such as median filtering or morphological operators. To this end, one might exploit some of the ideas in [60,65,54,43]. In this work, we consider only non-adaptive transforms. Our bias against adaptive transforms arises from simple practical considerations. Adaptive transforms are inherently more complex. Moreover, in the application of image coding (which is our focus herein), other difficulties arise when adaptive transforms are employed (e.g., how to syn-chronize the adaptation algorithm in the encoder and decoder). Although we do not explicitly consider the adaptive case in our work, one can certainly construct adaptive transforms using the GRITIT framework. That is, displace Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 44 operations can certainly employ adaptive filters. For example, one might exploit adaptive filtering by using some of the related ideas in [58,57]. Another slight variation on the GRITIT framework can be obtained by changing the displace operations to employ modular arithmetic. By using modular arithmetic, one can construct transforms that avoid dynamic range growth. Such an idea has been proposed, for example, by Chao et al. [44] in the context of the lifting framework. In the opinion of this author, however, such an approach is not particularly practical when lossy coding may be desired. If modular arithmetic is employed, the transform behavior can become extremely nonlinear. That is, small perturbations in the transform coefficients can result in very large changes in the reconstructed signal. Obviously, such behavior is undesirable in the case of lossy coding, since the effect of transform coefficient quantization (on distortion) can become quite unpredictable. When the GRITIT framework is used to construct reversible ITI wavelet transforms, the resulting computational structure is essentially a polyphase realization of a UMD filter bank. In such a scenario, we have a structure analagous to that in Figure 2.10. Such a structure is desirable from the standpoint of computational complexity, since analysis and synthesis filtering are performed in the downsampled domain (i.e., at the lower sampling density). We could, however, perform the analysis and synthesis filtering in the upsampled domain. That is, we could use a structure analagous to that in Figure 2.9. An approach like this has been proposed by Komatsu and Sezaki [88,89]. Clearly, from a computational standpoint, such an approach is less attractive (due to filtering operations being performed at the higher sampling density). For this very reason, we do not consider this type of scheme in detail in this thesis. We are interested exclusively in ITI transforms in this thesis. One can, however, construct (reversible) real-to-real transforms using the GRITIT framework. This is accomplished by modifying the displace operation such that the rounding operator quantizes with some granularity finer than integers. For example, one might employ an operator that rounds results to the nearest integer multiple of 2~1 0. In passing, we note that the GRITIT framework can also be employed to build reversible ITI transmultiplexors (i.e., the "dual" of analysis-synthesis filter banks). This possibility is not explored further in this thesis, however, as our application of interest (i.e., image coding) does not necessitate the use of transmultiplexors. 3.5 Relationship Between GRITIT and Other Frameworks As suggested earlier, the proposal of the GRITIT framework was partially motivated out of the desire to have a single unified view of frameworks for reversible ITI wavelet/block transforms. In the sections that follow, we examine the relationship between the GRITIT framework and each of a number of other frameworks, including the S+P transform, lifting, and ORT frameworks. As well, we consider the relationship between the GRITIT framework and the well known S transform. From this analysis, we are able to obtain some interesting insights into the interrelationships between the various frameworks. 3.5.1 S Transform First, we consider the relationship between the GRITIT framework and the S transform. Although the S transform is an example of a specific transform, rather than a transform framework, this relationship is still an interesting one to examine. The S transform was defined earlier in Section 3.3.1. By examining the network associated with the forward transform, shown in Figure 3.2(a), we can express the forward transform in operator notation (using the notation defined in Section 3.4.1) as L (0, L-J ,0, [o ±] ,0) £(1,0,0, [-1 o] ,0) £(0,1)2>(2, [o i]). (3.41) Similarly, the inverse transform, associated with the network shown in Figure 3.2(b), can be expressed in operator notation as .7(2,[o i])£(0,1)£(1,0, l,[- i o] ,0)L(0, L-J , 1, [o i] ,0). (3.42) From (3.41) and (3.42), we can see that the forward and inverse transforms are each comprised of a split/join, exchange, and two displace operations. Although the reversible nature of the S transform is not immediately obvious when the transform is viewed in its traditional form (i.e., as (3.32)), when viewed as a transform derived from the GRITIT framework, we can easily deduce that the S transform is reversible (and ITI). Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 45 3.5.2 S+P Transform Framework Now, let us consider the relationship between the GRITIT and S+P transform frameworks. The S+P transform frame-work was described earlier in Section 3.3.2, and is associated with the network shown in Figure 3.3. By examining the network in Figure 3.3(a), we can express the forward transform in operator notation as £(l,bfloor, 1, [A(z) -B(z)\,0)£ (0, L-J ,0, [o \] ,0) £(1,0,0, [-1 o] ,0) £(0,1)!P(2,[o i]). (3.43) Similarly, by examining the network in Figure 3.3(b), we can express the inverse transform as 3(2,[o i])-£(0,1)£(1,0,1,[-1 o] ,0)£(0, L-J, 1, [o\) ,0) £(l,bfloor,0,[A(Z) O],-B(Z)). (3.44) From (3.43) and (3.44), we can see that the forward and inverse transforms are each comprised of a split/join, exchange, and three displace operations. 3.5.3 Lifting Framework Next, we consider the relationship between the GRITIT and lifting frameworks. The lifting framework was described in Section 3.3.3, and is associated with the network shown in Figure 3.7. By examining the network in Figure 3.7(a), we can express the forward transform in operator notation as 5 p _ i - - - S i S 0 L x _ 1 - - - L 1 L 0 / 3 (3.45) where P is a split operation, the {£,} are scale operations, and the {L,} are displace operations. The scale operations {5,} are constrained to have nonzero integer scaling factors, in order for the resulting transform to be both reversible and ITI. The displace operations {L,} are constrained such that the AK(Z) and B ( z ) parameters in (3.36) are zero. By examining the network in Figure 3.7(b), we can express the inverse transform as From (3.45) and (3.46), we can see that the forward and inverse transforms are each comprised of one split/join operation, X displace operations, and p scale operations. 3.5.4 ORT Framework Now, we consider the relationship between the GRITIT and ORT frameworks. The ORT framework was described in Section 3.3.4, and is associated with the network shown in Figure 3.9, where polyphase filtering is accomplished using the networks illustrated in Figure 3.10. In what follows, we will show that the ORT networks in Figures 3.10 can be realized using GRITIT networks. Consider the networks shown in Figures 3.15(a), 3.15(b), 3.15(c), 3.15(d), 3.15(e), 3.15(f), and 3.15(g). In operator notation, these networks can be expressed, respectively, as 5 ( l , -* )£ (0 , L-J ,0, [o \-H(Z)] ,O)£(1,0,O, [-1 o] ,0), (3.47a) £(1,0,1, [-1 o] ,0)£(0, L-J, 1, [o 1-ffM] ,0)5(1, - 1 / s ) , (3.47b) 5(l,±l)£(l ,0,O,[±io],O)£(O, L-J ,0,[o H(z)},0), (3.47c) £(0, L-J, 1, [o H(z)} ,O),£(1,0,1, [±i 0] ,0)5(1,±1), (3.47d) 5(1,*)£(0,L-J,0,[0H(Z)],0), (3-47e) £(0, L-J, l,[o»W], 0)5(1,1 A), and (3.470 5(1,-1)£(O,0,O,[o i] ,O)£(1,0,O, [-1 o] ,O)£(O,0,O,[o i] ,0). (3.47g) Clearly, these networks are comprised of only displace and scale operations, and therefore, can be constructed using the GRITIT framework. Now, we will prove that the ORT networks depicted in Figures 3.10(a), 3.10(c), 3.10(e), and 3.10(h) are equivalent to the GRITIT networks shown in Figures 3.15(a), 3.1.5(c), 3.15(e), and 3.15(g), respec-tively. Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 46 *o[n] X L-J yo[n] yo["] L-J' x\[n\ L-J H(z) \-H(z) (a) ±1 -<•>• (c) yil"] yo[n] ± i yoM L J H{z) xx [n] (e) yi[«] (b) ±1 L-J H(z) xi[n] yiw ± i jri[n] (d) yoM L-J H{z) *o[n] xm (f) Figure 3.15: Equivalent GRITIT networks for the (a) A-type, (b) A '-type, (c) B-type, (d) B '-type, (e) C-type, (f) C~'-type, and (g) E-type networks of the ORT framework. Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 47 First, let us consider the A-type network shown in Figure 3.10(a). By observing that xo[n] and xi[n] are integer sequences and using (2.1), this network can be shown to be equivalent to that in Figure 3.15(a). In the case of the network in Figure 3.10(a), the relationship between the inputs and outputs is given by Y0(z) = [H(z)X0(z) + (1 -tf(z))Xi(z)Jz = X l ( z ) + [ H ( z ) ( X 0 ( z ) - X i ( z ) ) \ t , Yl(z) = s ( X 0 ( z ) - X l ( z ) ) . and Similarly, for the network in Figure 3.15(a), the input-output behavior is given by Y o ( z ) = X 0 ( z ) + [ ( l - H ( z ) ) ( X l ( z ) - X 0 ( z ) ) \ z = X0(z) + [ X l ( z ) - X o ( z ) + H ( z ) ( X 0 ( z ) - X i ( z ) ) \ t = X 1 ( z ) + \ . H ( z ) ( X 0 ( z ) - X i ( z ) ) \ z , and Y1(z) = - s ( X l ( z ) - X 0 ( z ) ) = s ( X 0 { z ) - X l ( z ) ) . (3.48a) (3.48b) (3.49a) (3.49b) By comparing (3.48) and (3.49), it is clear that both networks have the same input-output characteristics. In other words, we have shown that the structures in Figures 3.10(a) and 3.15(a) are equivalent. Thus, an A-type network can be realized using only displace and scale operations (as given by (3.47a)). Next, we consider the B-type network shown in Figure 3.10(c). Using the same approach as above, this structure can be shown to be equivalent to that in Figure 3.15(c). In the case of the network in Figure 3.10(c), the relationship between the inputs and outputs is given by Yo(z) = [ X 0 ( z ) + H ( z ) X l ( z ) \ z (3.50a) = X0(z) + [ H { z ) X 1 ( z ) \ z , and Y\ (z) = L*o(z) + (#(z) ± 0*1 ( z ) \ z (3.50b) = [ X 0 ( z ) ± X l ( z ) + H ( z ) X l ( z ) \ z = X 0 ( z ) ± X l ( z ) + [ H ( z ) X i ( z ) \ z . Similarly, for the network in Figure 3.15(c), the input-output behavior is characterized by Y o { z ) = X 0 ( z ) + l H { z ) X 1 { z ) \ z , and (3.51a) Yi(z) = ± ( X l { z ) ± Y 0 ( z ) ) (3.51b) = ± ( X 1 ( z ) ± ( X 0 ( z ) + l H ( z ) X 1 ( z ) \ z ) ) = ± X i ( z ) + X 0 ( z ) - r l H { z ) X l ( z ) \ z = X 0 ( z ) ± X i ( z ) + [ H ( z ) X i ( z ) \ z . Comparing (3.50) and (3.51), we see that both networks have the same input-output characteristics. In other words, the filtering structures in Figures 3.10(c) and 3.15(c) are equivalent. Thus, a B-type network can also be realized using only displace and scale operations (as given by (3.47c)). Next, let us consider the C-type network shown in Figure 3.10(e). This is equivalent to the network shown in Figure 3.15(e). To see this, we simply invoke identity (2.1). Hence, a C-type network can be realized using only displace and scale operations (as given by (3.47e)). Lastly, consider the E-type network shown in Figure 3.10(h). This linear system has the transfer matrix T = 0 1 1 0 The network shown in Figure 3.15(g) has the transfer matrix 0 1 1 1 0 1 1 0 1 0 1 -1 1 0 1 1 0 (3.52) (3.53) Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 48 Comparing (3.52) and (3.53), we see that both networks have the same input-output behavior. In other words, the structures in Figures 3.10(h) and 3.15(g) are equivalent. Thus, an E-type network can be realized using only displace and scale operations (as given by (3.47g)). Using techniques similar to those employed above, the networks in Figures 3.10(b), 3.10(d), and 3.10(f) can also be shown to be equivalent to those in Figures 3.15(b), 3.15(d), and 3.15(f), respectively. The only additional tools needed are identities (2.2) and (2.3). For example, consider the networks of Figures 3.10(b) and 3.15(b). We proceed by using identities (2.1), (2.2), and (2.3), and the fact that j - 1yi[n] € Z. For the network in Figure 3.10(b), the relationship between the inputs and outputs is given by X o ( z ) = \ Y 0 ( z ) + s - l ( ( l - H ( z m ( z ) ] z (3.54a) = Y0(z) + \ s - l Y t ( z ) - s - 1 H ( z ) Y l (z)\ = Y 0 ( z ) + s - l Y l ( z ) + \ - s - l H ( z ) Y i ( z ) ] z = Y 0 ( z ) + s - 1 Y i ( z ) - [ s - l H ( z ) Y i ( z ) \ t , and Xi{z) = \ Y 0 ( z ) - s - l H ( z ) Y l ( z ) ] z (3.54b) = Y 0 ( z ) + \ - s - 1 H ( z ) Y i ( z ) \ = Y 0 ( z ) - [ s - l H ( z ) Y l ( z ) \ z . In the case of the network in Figure 3.15(b), the inputs and outputs are related by X0(z) = Y0(z) - [ s - l ( l - H ( z ) ) Y l ( z ) \ z (3.55a) = Y o ( z ) - [ - s - l Y i ( z ) + s - 1 H { z ) Y l ( z ) \ z = Y0(z) + s - l Y l ( z ) - [ s - l H { z ) Y i ( z ) \ t , and X l ( z ) = X 0 ( z ) - s - 1 Y l ( z ) (3.55b) = Y 0 ( z ) - [ s - i H ( z ) Y l ( z ) \ z . Comparing (3.54) and (3.55), we see that the two networks have the same input-output behavior. In other words, the networks in Figures 3.10(b) and 3.15(b) are equivalent. This result should not be surprising, however. In fact, it must be so. If we have two invertible transformations with identical forward transforms, the inverse transforms must also be identical. This argument can be used to show that the remaining networks are also equivalent. To summarize our results so far, we have shown that the A-, A - 1 - , B-, B - 1 - , C-, C - 1 - , and E-type networks of the ORT framework can be realized using the GRITIT networks (as given by (3.47) and Figure 3.15). Next, we observe that the D-type ORT network is exactly equivalent to the GRITIT shift operation. Consequently, the polyphase filtering networks employed by the ORT framework can be realized using GRITIT networks. Thus, the ORT framework is simply a special case of the GRITIT framework. More specifically, with the ORT framework, the forward transform has the form Ts-i—nToP where P is a split operation and the {7;} are chosen from shift operations or operations of the forms given in (3.47a), (3.47c), (3.47e), and (3.47g). The inverse transform follows immediately from the forward transform: 3.6 Relationship Between the ORT and Lifting Frameworks Although the ORT and lifting frameworks appear quite different at first glance, they are, in fact, intimately related. In what follows, we will show that the ORT framework is, in fact, a special case of the lifting framework with only trivial extensions. In Section 3.5.4, we showed that ORT polyphase filtering networks can be decomposed into GRITIT networks as given by (3.47). These GRITIT networks, however, only utilize displace and scale operations that have the same form Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 49 mm ui[n] «0|"J 7* v0[n] «oH + ) » > 1 + L-J H(z) z~KH{z) vi[n] (a) _ voW uo[n] ^ L-J L-J H(z) J | z*H(z) vi [n] «l [fi] (b) vo["J v,[n] vol"J zK vi [n] v,n u.-n Vj[fl] (C) Figure 3.16: Identities for interchanging (a) D-type and L-type operations (first case), (b) D-type and L-type operations (second case), and (c) D-type and S-type operations. as the L- and S-type operations from the lifting framework. Therefore, the ORT networks can also be decomposed into L- and S-type operations. In the ORT framework, the analysis polyphase filtering is performed by a cascade of A-, B-, C-, D-, and E-type networks in any number and any order. Since the A-, B-, C-, and E-type networks can be realized using only L- and S-type networks, it follows that such an analysis polyphase filtering network can be replaced by an equivalent one consisting of only L-, S-, and D-type operations in any number and any order. This new filtering structure, however, is exactly that employed by the lifting framework with two simple extensions: 1) the L- and S-type operations are allowed to be intermixed, and 2) the use of D-type operations is also permitted. One might wonder what additional degrees of freedom are offered by these extensions. As we shall demonstrate, these extensions are, in fact, of very little practical value. Initially, let us consider the second extension above. Suppose we allow D-type operations to be used. One can easily show that, given a network where D-type operations are intermixed with other operations, it is always possible to find an equivalent network with the D-type operations applied last. Using (2.15) and straightforward manipulation, one can easily confirm that the identities shown in Figure 3.16 hold. By repeated application of these identities, one can always obtain a new network with the D-type operations applied last. From this, we can see that there is little flexibility afforded by the second extension listed above. This extension only allows the introduction of an additional shift in the subband signals. Clearly, this does not change the underlying transform in any fundamental way. Let us now consider the first extension above. For lossless signal coding applications, we would like a transform that yields coefficients having small magnitudes. Such coefficients are desirable as they can be more efficiently en-coded. This being the case, it is undesirable to employ S-type operations having scaling factors other than ±1. The use of such operations would unnecessarily increase the dynamic range of the subband signals and, hence, the resulting transform coefficients as well. Suppose now that we restrict the scaling factors to be ±1 . Since scaling by one has no effect, we need only consider the case of scaling by — 1. In this scenario, one can easily show that given a network of L- and S-type operations where the two types of operations are intermixed, it is always possible to find an equivalent network with the S-type operations applied last. Using (2.3) and straightforward manipulation, we can show that the Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 50 71 X - N vo[n] «o[n] ^ -1 H(z) ui[n] -H(z) n[n] ui[n] (a) vi [n] uo[n\ vo[«] «o[n] vol«J H(z) - l vi [n\ «i [nj -1 vi [n] (b) Figure 3.17: Identities for interchanging L-type and S-type operations (in the scaling by -1 case). identities in Figure 3.17 hold. By repeated application of these identities, one can always obtain a new network with the S-type operations applied last. Although the L-type operations in the resulting network may employ either the floor or ceiling function for rounding, the network is still a lifting network. As noted previously, the lifting framework allows any rounding operation to be used in the L-type networks. As we can see from the above analysis, allowing a scaling factor of — 1 only allows one to change the sign of subband signals. Thus, there is little to be gained by employing S-type operations in the first place. As we have shown above, the additional degrees of freedom offered by the two lifting framework extensions are of little practical value. Thus, we assert that there are no practically useful transforms that can be generated with the ORT framework that cannot be generated with the lifting framework. As a final note, it is important to emphasize the generality of lifting. For the purposes of showing the relationship between the ORT and lifting frameworks, it was necessary to use the floor and ceiling functions for the rounding operations in the lifting case. In general, however, the lifting framework allows the use of an arbitrary rounding function (e.g., floor, ceiling, truncation, rounding to the nearest integer, etc.). In fact, with the appropriate choice of rounding function, the lifting framework can even be used to generate reversible real-to-real mappings [9]. This, however, is not generally possible with the ORT framework due to its use of the A- and B-type networks. The lifting framework has another important advantage over the one based on the ORT. Lifting always facilitates in-place calculation of a transform. This, however, is not generally true for the ORT framework. This is, again, due to its use of the A- and B-type networks. Therefore, lifting is arguably a more powerful and practical framework than that based on the ORT. 3.7 Generalized S Transform In Section 3.4.3, we explained how the GRITIT framework can be used to construct reversible ITI block transforms. Here, we provide a concrete example to demonstrate the applicability of the GRITIT framework in this context. In the sections that follow, we propose the generalized S transform (GST), a family of reversible ITI block transforms based on the key ideas behind the S transform. We then study the GST in some detail. This leads to a number of interesting insights about transforms belonging to the GST family, the S transform, and reversible ITI transforms in general. The remainder of this material is structured as follows. The block S transform is introduced in Section 3.7.1, and the GST family of transforms is defined in Section 3.7.2. Sections 3.7.3 and 3.7.4 proceed to examine GST parameter calculation and the effects of using different rounding operators in the GST. Some examples of well known transforms belonging to the GST family are given in Section 3.7.5, and in Section 3.7.6, we present a new GST-based transform, Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 51 and demonstrate its utility for image coding applications. 3.7.1 Block S Transform In Section 3.3.1, we introduced the S transform, one of the simplest and most ubiquitous reversible ITI wavelet transforms. The S transform is associated with the networks shown earlier in Figure 3.2. By examining these networks, we can see that the polyphase filtering operations of the S transform themselves constitute a simple two-point block transform, which we refer to as the block S transform. That is, the block S transform is associated with the networks shown in Figure 3.18, where the networks in Figures 3.18(a) and 3.18(b) correspond to the forward and inverse transforms, respectively. Mathematically, the forward transform maps the integer vector [*o *i ] r to the integer vector [yo yi]T, as given by xo-xi J The corresponding inverse transform is given by [*?] = [,_!„], where t =\ y 0 + [ f a i + 1)J. (3.56) 3.7.2 Generalized S Transform By examining the block S transform in the context of the GRITIT framework, we are inspired to propose a natural extension to this transform, which we call the generalized S transform (GST). For convenience in what follows, let us define two integer vectors x and y as x = [*o x\ ... XN-I] and y = [yo yi ••• y/v-i J . The forward GST is a mapping from x to y of the form y = Cx + Q{{B-I)Cx), (3.57) where B is real matrix of the form • l>N-l' 0 1 0 . . 0 0 0 1 . 0 -6 6 6 . i -C is a unimodular (i.e., |detC| — 1) integer matrix defined as C0,N-1 C2,N-1 c0,0 C0,1 £0,2 cl,0 C2,0 c2,l Q,2 c/v-i.o cyv-i,i CN-l,2 ••• Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 52 (a) (b) Figure 3.19: Network realization of the generalized S transform, (a) Forward transform and (b) inverse transform. and Q is a rounding operator. By examining (3.57), one can easily see that this transform maps integers to integers. In the absence of the rounding operator Q, the GST simply degenerates into a linear transform with transfer matrix A, where A = BC. Thus, the GST can be viewed as a reversible ITI mapping that approximates the linear transform characterized by the matrix A. The inverse GST is given by x = C-l(y-Q((B-I)y)). (3.58) Note that C~l is an integer matrix, since, by assumption, C is a unimodular integer matrix. To show that (3.58) is, in fact, the inverse of (3.57), one need only observe that due to the form of B, for any two N x I vectors u and v. v = u + Q((B-I)u) (3.59) implies u = v-Q((B-I)v). (3.60) By substituting u = Cx and v = y into (3.59) and (3.60), we obtain (3.57) and (3.58), respectively. As suggested earlier, the GST is an example of a reversible ITI block transform generated using the GRITIT framework. The GST can be realized using the networks shown in Figure 3.19. The forward transform, associated with the network in Figure 3.19(a), is comprised of a unimodular integer transfer matrix C followed by a single GRITIT displace operation (i.e., L (0,(2,0, [o ft, b2 ••• ,0)). Furthermore, the unimodular integer matrix C can always be realized using the exchange, scale, and (linear) displace operations from the GRITIT framework. To see why this is so, we observe that any unimodular integer matrix can be decomposed into a product of elementary unimodular integer matrices. (The proof that a unimodular integer matrix can be decomposed into a product of elementary unimodular integer matrices is constructive by the standard algorithm for computing the Smith normal form of an integer matrix (e.g., as described in [56]).) Since any elementary unimodular integer transfer matrix trivially has a corresponding GRITIT operation (i.e., a displace, scale, or exchange operation), we can always find a GRITIT-based realization of a unimodular integer transfer matrix. Therefore, the GST is clearly generated by the GRITIT framework. From a mathematical viewpoint, the specific strategy used to realize the transforms C and C _ 1 is not particularly critical. Since both transforms are linear, each has many equivalent realizations. These two transforms could be implemented using GRITIT networks, as suggested above, but this is not necessary. In passing, we would like to note that Hao and Shi [63] have done some related work with /V-point block transforms. Their work is concerned primarily with the lifting factorization problem for general block transforms, whereas we study in detail a very specific class of these transforms (i.e., those belonging to the GST family). Thus, the results presented herein and those presented in [63] are complementary. Other works, such as [36,1 ], have also considered reversible N-inputAf-output ladder networks in various contexts (where N > 2). These other works, however, do not specifically consider the GST family of transforms. Thus, our results are distinct from those in the preceding works. 3.7.3 Calculation of GST Parameters Suppose that we are given a linear transform characterized by the transform matrix A, and we wish to find a reversible ITI approximation to this transform based on the GST framework. In order for this to be possible, we must be able to Chapter 3. Frameworks for Reversible ITI Wavelet/Block Transforms 53 decompose A as A = BC (3.61) where the matrices B and C are of the forms specified in (3.57). Therefore, we wish to know which matrices have such a factorization. The answer to this question is given by the theorem below. Theorem 1 (Existence of G S T Factorization). The real matrix A, defined as "0,0 "0,1 "1,0 « l , l " O . A ' - l <*l,N-\ _aw-1,0 aw-1,1 ••• "N-ljl-\_ has a GST factorization (i.e., a factorization of the form of(3.61)) if and only if all of the following conditions hold: 1. The last N—l rows of A must contain only integer entries. That is, a i J € Z for i=\,2,...,N-I, j = 0,1,...,N-\. 2. A is unimodular. 3. The integers {det minor (.,4,0, OJM)1 a r e relatively prime. Proof. First, we prove the sufficiency of the above conditions. We begin by considering the slightly more general decomposition A = BDC (3.62) where B and C are defined as in (3.61), and D A b0 0 0 ... 0-0 1 0 ... 0 0 0 1 ... 0 •0 00 0 1-A comparison of the left- and right-hand sides of (3.62) yields the trivial relationships cij=aij fori= l , 2 , . . . , t f - l , ; = 0 , l , . . . , t f - l and the nontrivial system of equations a = bC (3.63) where a = [ f lo,o ao,i - a o ,A f - i ] and b = [bo h - <w- i ] . If C is nonsingular, we can solve for b in terms of a as b = aC By considering the determinants of the various matrices in the factorization, we can write detvl = det(BDC) = (detB)(det£>)(detC) = b0detC. Therefore, bo = (det>l)/(detC). Consequently, we want to choose C such that detC = detA. In this case, b0 — 1, so D = I, and we have A = BDC = BC, a factorization of the desired form. Let w,- = detminor(A,0,i) for i = 0,l,...,N—l. From (3.63), we know wt ==detminor(C,0,i) (3.64) for i = 0,1,..., AT — 1. Using the Laplacian expansion for the determinant of C across row 0, we obtain detC = S^- 0 1(-l)'co,,w ;. (3.65) Chapter 3. Frameworks for Reversible ITI Wavelet/Block Transforms 54 Since det A G {-1,1} and we want to choose the {co,/}^1 such that detC = det A, from ( 3 . 6 5 ) we must solve where c'0i = co,i/detA and w't = (-l)'tv,- for i = 0,1,...,N — 1. In this equation, the {wj}^' 3 X 6 relatively prime since the {w,}^ 1 are (by assumption) relatively prime, and the {c'0 J ^ L Q 1 a r e integers since the {co,,}^1 are integers and det A G {—1,1}. We can show that ( 3 . 6 6 ) always has a solution. This result follows immediately from the well known result in number theory that states: The greatest common divisor (GCD) of a set of integers is a linear combination (with integer coefficients) of those integers, and moreover, it is the smallest positive linear combination of those integers [ 1 2 1 ] . In our case, the integers {w-}^ 1 are relatively prime, so their G C D is 1, and therefore, 1 must be a linear combination of the integers {w-}^ 1. Furthermore, the coefficients of the linear combination can be found using the Euclidean algorithm [ 1 2 1 ] . Therefore, we can use ( 3 . 6 6 ) to generate a valid choice for C, and knowing C, we can solve for b, and hence find B. Thus, we have a constructive proof that the desired factorization of A exists. This shows the sufficiency of the above conditions on the matrix A. The necessity of the above conditions on the matrix A is easily shown. Due to the form of B, the last N — 1 rows of the matrices A and C must be the same. Thus, the last N— 1 rows of A must contain only integer entries, since C is an integer matrix. If A is not unimodular, obviously it cannot be decomposed into a product of two unimodular matrices (namely, B and C). This follows from the fact that detA = (detB)(detC), detB G {-1,1}, and detC e {-1,1}. The relative primeness of {detminor(.A,0,t)}^o1 follows from the " G C D is a linear combination" theorem stated above. If this set of integers is not relatively prime, their G C D is greater than one, and no solution to ( 3 . 6 6 ) can exist. Thus, the stated conditions on the matrix A are necessary for a GST factorization to exist. • From the proof of the above theorem, we know that the factorization is not necessarily unique, since more than one choice may exist for the {crj,;}^1 above. In instances where the solution is not unique, we can exploit this degree of freedom in order to minimize the computational complexity of the resulting transform realization. For most practically useful transforms in the GST family, it is often the case that {minor(A,0,i)}^="01 are all unimodular (i.e., |detminor(j4,0,i)| = 1 for all i G {0, \ ,...,N — 1}). This condition holds for all of the examples considered later in Section 3 . 7 . 5 . In such instances, the most useful GST factorizations are often obtained by choosing the unknowns {co,,}^1 so that all but one are zero, with the remaining unknown being chosen as either —1 or 1 in order to satisfy ( 3 . 6 6 ) . In this way, we can readily obtain N distinct GST factorizations of A which typically yield low complexity transform realizations. In all likelihood, the most practically useful transforms in the GST family are those for which the output yo is chosen to be a rounded weighted average of the inputs {xi}1^ (with the weights summing to one) and the remaining outputs {y,}^ 1 are chosen to be any linearly independent set of differences formed from the inputs {xflftj^. Such transforms are particularly useful when the inputs {xi}^T0l tend to be highly correlated. Obviously, there are many ways in which the above set of differences can be chosen. Here, we note that two specific choices facilitate a particularly computationally efficient and highly regular structure for the implementation of C (and C - 1 ) . Suppose the difference outputs are selected in one of two ways: typel: yt =xt -x0 type 2: y; = -x,-\ for i = 1,..., N — 1. Since these differences completely define the last N-l rows of A, we can calculate det minor( A, 0,0). Furthermore, one can easily verify that in both cases detminor(.A,0,0) = 1. Thus, we may choose C for type-1 and type-2 systems, respectively, as r l C O . • 0 i r l 0 o .. . 0 on - 1 1 0 . . 0 - l l o .. . 0 0 - 1 0 1 . . 0 and 0 - l I .. . 0 0 • 0 • 0 0 -—i 6 6 . . lJ - 6 6 6 .. . - l iJ Furthermore, we can show that the corresponding solutions for B are, respectively, given by typel: 6, = ao,< type 2: bt = j j ^ 1 ao* Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 55 VQ "2. " N - J ? N - \ (a) v 0 "2^ UN VN-2 (c) vn «o VI V2 V2 V N -vn " 1 "2 KJV-1 (b) «o 6 -" l » 2 VJV-2 " N - l (d) Figure 3.20: Networks for realizing particular forms of the C and C 1 matrices. The networks for type-1 systems: (a) forward and (b) inverse networks. The networks for type-2 systems: (c) forward and (d) inverse networks. for i = 1,2,...,N-l. In the above cases, the transform matrices for C and C - 1 can be realized using the ladder networks shown in Figure 3.20. The forward and inverse networks for a type-1 system are shown in Figures 3.20(a) and (b), respectively. Similarly, the forward and inverse networks for a type-2 system are shown in Figures 3.20(c) and (d). For each of the two system types, the forward and inverse networks have the same computational complexity (in terms of the number of arithmetic operations required). Other choices of differences also facilitate computationally efficient implementations. Choices that yield a C matrix with all ones on the diagonal are often good in this regard. 3.7.4 Choice of Rounding Operator So far, we have made only one very mild assumption about the rounding operator Q (i.e., property (2.12) holds). At this point, we now consider the consequences of a further restriction. Suppose that the rounding operator Q is integer-bias invariant (as defined by (2.13)). The operator Q could be chosen, for example, as the floor, biased floor, ceiling, or biased ceiling function (as defined in Section 2.2). Note that the truncation and biased truncation operators are not integer-bias invariant. To demonstrate this for either operator, one can simply evaluate the expressions on the left- and right-hand sides of equation (2.13) for a = —^ and x = 1. Clearly, the two expressions are not equal (i.e., 0 ^ 1 or 1^0). If Q is integer-bias invariant, we trivially have the two identities shown earlier in Figure 3.1. Therefore, in the case that Q is integer-bias invariant, we can redraw each of the networks shown in Figures 3.19(a) and 3.19(b) with the rounding unit moved from the input side to the output side of the adder. This results in the networks shown in Figures 3.21(a) and 3.21(b). Mathematically, we can rewrite (3.57) and (3.58), respectively, as y = Cx + Q((B-I)Cx) = Q(Cx + (B-I)Cx) = Q(BCx), and (3.67a) Chapter 3. Frameworks for Reversible ITI Wavelet/Block Transforms 56 *2_l XN-l , yo \ * " • -1 b i \ \ b 2 bN-\ (a) yo yi yi HeH-i yjv-j b i \ \ b 2 bN-i ( *0 ( *1 W-l (b) Figure 3.21: Equivalent representation of the generalized S transform for an integer-bias invariant rounding operator, (a) Forward transform and (b) inverse transform. x Or alternately, in the latter case, we can write where = C-i(y-Q((B-I)y)) = -C-1(Q((B-I)y)-y) = -C-xQ{{B-I)y-y) = -C-xQ(.{B-2I)y) = -C-lQ(-B-ly). x = C'lQ'(B-ly) Of (a) = - G ( - a ) . (3.67b) At this point, we note that Q and Q' must be distinct (i.e., different) operators. This follows from the fact that a rounding operator cannot both be integer-bias invariant and satisfy 2(a) = -Q(-ct) for all a e i . (See Proposition 1 for a simple proof.) By using (2.3), we have, for the case of the floor and ceiling functions W for Q(x) = [x\ [x\ for fi(*) =r*l-The above results are particularly interesting. Recall that the GST approximates a linear transform characterized by the matrix A = BC. Examining the definition of the forward GST given by (3.57), we can see that, due to the Chapter 3. Frameworks for Reversible ITI Wavelet/Block Transforms 57 rounding operator Q, the underlying transform is a function of both B and C individually, and not just their product A. In other words, different factorizations of A (into B and C) do not necessarily yield equivalent transforms. When Q is integer-bias invariant, however, (3.57) simplifies to (3.67a), in which case the underlying transform depends only on the product A. Thus, for a fixed choice of integer-bias invariant operator Q, all GST-based approximations to a given linear transform (characterized by A) are exactly equivalent. This helps to explain why so many equivalent realizations of the S transform are possible, as this transform utilizes the floor function which is integer-bias invariant. 3.7.5 Examples One member of the GST family is, of course, the block S transform. We can factor the transform matrix A as A = BC where • t 1 ' A = 2 2 , B = . 1 -1 . Lo I J and C ro l , i - l The block S transform is obtained by using the parameters from the above factorization for A in the GST network in conjunction with the rounding operator Q(x) = [x\. (The particular network used to realize the matrix C is indicated by the factorization of C shown above.) Mathematically, this gives us where t = XQ — x\ and = [»+'] where 5 = y 0 - ^ J -(3.68a) (3.68b) Comparing (3.56) to (3.68b), we observe that the computational complexity of the latter expression is lower (i.e., one less addition is required). Due to our previous results, however, we know that both equations are mathematically equivalent. Thus, we have found a lower complexity implementation of the inverse block S transform. In turn, this also yields a reduced complexity implementation of the inverse S transform (which employs the inverse block S transform for polyphase filtering). Another example of a transform from the GST family is the reversible color transform (RCT), defined in the JPEG-2000 Part-1 standard [72] (and differing only in minor details from a transform described in [59]). Again, we can factor the transform matrix A as A = BC where r 1 1 1 "I r. 1 1 • 4 2 4 , B = 1 4 4 0 -1 1 0 1 0 L 1 -1 oJ Lo o 1 J and C 1 01 r 0 -1 1 . i -1 0. 1 00 -1 1 0 0 oi r 0 l J . 1 00 0 1 0 -1 0 1 ro l oi 0 0 1 .1 0 0. The RCT is obtained by using the parameters from the above factorization for A in the GST network in conjunction with the rounding operator Q(x) = \x\ • (Again, the particular network used to realize the matrix C is indicated by the factorization of C shown above.) Mathematically, this gives us yo' y\ lyi. *o *\ x2 xi+[\('o+h) ]-[ 'o s s+yi where '?=%ZXX\ and The formula given for the forward RCT in [ 7 2 ] is wheres = y0- [\(yi +yi)\ • X 2 - X \ xo-xi (3.69a) (3.69b) (3.70) By comparing (3.69a) and (3.70), we observe that the computational complexity of the former expression is lower (i.e., four adds and one shift are required instead of, say, four adds and two shifts). Thus, we have found a lower complexity implementation of the forward RCT. Although the computational complexity is reduced by only one operation, this savings is very significant in relative terms (since only six operations were required before the reduction). Recall that for integer-bias invariant rounding operators, multiple realization strategies often exist for a particular GST-based reversible ITI transform. In order to demonstrate this, we now derive an alternative implementation of the Chapter 3. Frameworks for Reversible ITI Wavelet/Block Transforms 58 RCT. To do this, we factor the transform matrix associated with the RCT (i.e., the matrix A from above) as A -where BC 1 =1 I 1 4 4 0 1 0 0 and C = ro o li ri o oi r l ooi ri o o- ro o n 0 -1 1 0 - 1 0 -1 1 0 0 1 0 0 1 0 . 1 - 1 0 . Lo o iJ L o o l J .0 -1 lj .1 o oJ The alternative RCT implementation is obtained by using the parameters from the above factorization for A in the GST network in conjunction with the rounding operator Q(x) = [x\. (Again, the particular network used to realize the matrix C is indicated by the factorization of C shown above.) The corresponding RCT implementation is given by -c2+|_5(-3'o+'i)J '0 '1 = s0 . *1 where where J£g:*; and si=yo-[j(-3yi+y2)\' One can see that the computational complexity of this alternative implementation is higher than the one proposed in (3.69). In fact, it is likely that, due to the simple nature of the RCT, the implementation given by (3.69) is probably the most efficient. 3.7.6 Practical Application: Modified Reversible Color Transform When compressing a color image, a multicomponent transform is often applied to the color components in order to improve coding efficiency. Such a transform is frequently defined so that one of its output components approximates luminance. In this way, one can easily extract a grayscale version of a coded image if so desired. For RGB color imagery, luminance is related to the color component values by y = 0.299r + 0.587g + 0.114b (3.72) where y is the luminance value, and r, g, and b are the red, green, and blue component values, respectively. The forward RCT is defined in such a way that one of its output components approximates luminance. In effect, the RCT computes the luminance as \r+^g+ jb. Obviously, this is a rather crude approximation to the true luminance as given by (3.72). Without a huge sacrifice in complexity, however, we can define a new RCT-like transform that provides a much better approximation to true luminance. To this end, we propose a slightly altered version of the RCT called the modified RCT (MRCT). The forward and inverse equations for the MRCT are given, respectively, by yo y\ In to n where J j z ^ and *0 " s0 -1 Xl si xi. .s\+y2. where 1_ 128 yo-\-m^+^yi) (3.73a) (3.73b) The MRCT is associated with the transform matrix A = 38 75 15 128 128 128 -1 1 0 0 -1 1 which can be factored as A = BC, where l 90 15 128 128 0 1 0 L0 0 1 and C = r i o oi r _ i i o . 0 -1 lJ L 1 0 0 -1 1 0 0 °i r1 0 0 0 1 iJ Lo - i The MRCT is obtained by using the above factorization for A in the GST network in conjunction with the rounding operator Q(x) = [x\. (The particular network used to realize the matrix C is indicated by the factorization of C shown above.) The forward and inverse MRCT can each be implemented with six adds and four shifts (where expressions of the form [T|g(90w + 15v)J can be computed as asr((16-l)[(4 + 2)w +v], 7)). Recall that the forward and inverse Chapter 3. Frameworks for Reversible ITI Wavelet/Block Transforms 59 Table 3.3: Deviation from the true luminance for the M R C T and R C T Image MRCT RCT MAE PAE MAE PAE f lov/ers 0.654 1.635 6.515 25.326 food 0.734 1.777 8.860 34.821 sailboat 0.566 1.438 1.888 20.580 tile 0.776 1.369 9.591 14.560 RCT, as given by (3.69), each require four adds and one shift. Thus, in terms of computational complexity, the M R C T compares quite favorably with the RCT, when one considers the RCT's extreme simplicity. Having defined the MRCT, we first studied its coding performance. A set of 196 RGB color images, covering a wide variety of content, was compressed losslessly using both the M R C T and R C T in the same coding system. The difference in the overall bit rate obtained with the two transforms was less than 0.1%. Thus, the M R C T and R C T are quite comparable in terms of coding efficiency. In addition to coding efficiency, we also examined the luminance approximation quality. For each image, we computed the true luminance component as given by (3.72) and the luminance components generated by the M R C T and RCT. We then proceeded to calculate the approximation error for each transform. The results for several of the test images are shown in Table 3.3. Clearly, the M R C T approximates true luminance much more accurately than the R C T (in terms of both M A E and PAE). Furthermore, this difference in approximation quality is usually visually perceptible. That is, the M R C T yields a luminance component that appears visually identical to the true luminance component, while the R C T generates a luminance component that is often visibly different from the ideal. When designing the MRCT, we had to choose which set of differences to employ (for the two non-luminance outputs of the forward transform). In passing, we would like to note that this choice can impact coding efficiency. For natural imagery, experimental results suggest that the difference between the red and blue components is often slightly more difficult to code than the other two differences (i.e., green-blue and green-red). For this reason, we chose to use the green-blue and green-red differences to form the non-luminance outputs of the (forward) M R C T . 3.8 Summary In this chapter, we proposed the generalized reversible ITI transform (GRITIT) framework, a single unified framework for reversible ITI wavelet/block transforms. This new framework was then used to study several previously developed frameworks and their interrelationships. For example, the ORT framework was shown to be equivalent to a special case of the lifting framework with only trivial extensions. Moreover, we demonstrated that any practically useful reversible ITI mapping that can be realized using the ORT framework can also be realized using the lifting one. This development is significant in that it demonstrates the generality of the lifting framework for the construction of reversible ITI transforms. Rounding operators were examined in some detail. The significance of the integer-bias invariance and oddness properties was discussed, and the mutual exclusivity of these two properties was demonstrated. Several common rounding operators were characterized in terms of their approximation properties (i.e., error interval, PAE, M A E ) and computational complexity. Relationships between these rounding operators were also examined. When considering both computational complexity and rounding error characteristics together, the floor, biased floor, and biased ceiling operators are arguably the most practically useful of those considered. The floor operator is desirable as it has the lowest computational complexity, while the biased floor and biased ceiling operators are slightly more computationally complex but normally have the best rounding error performance. Under certain circumstances, the floor function was shown to share the best rounding error performance, in spite of its low computational complexity, a fact that is often overlooked in the existing literature. The GRITIT framework was demonstrated (by example) to be useful for the construction of reversible ITI block transforms. The generalized S transform (GST), a family of reversible ITI block transforms, was proposed. Then, the GST was studied in some detail, leading to a number of interesting results. We proved that for a fixed choice of integer-bias invariant rounding operator, all GST-based approximations to a given linear transform are equivalent. The R C T was shown to be a specific instance of the GST. We showed how various degrees of freedom in the choice of GST parameters can be exploited in order to find transform implementations with minimal computational complexity. Chapter 3. Frameworks for Reversible I T I Wavelet/Block Transforms 60 Moreover, the GST framework was used as a means to derive reduced complexity implementations of the S transform and RCT. A new transform belonging to the GST family, called the MRCT, was introduced, and shown to be practically useful for image coding applications. Due to its utility, the GST family of transforms will no doubt continue to prove useful in both present and future signal coding applications. When I despair, I remember that all through history, the way of truth and love has always won. There have been murderers and tyrants, and for a time they can seem invincible. But in the end they always fall. Think of it, always. —Mohandas Karamchand Gandhi (1869-1948) 61 Chapter 4 Nonexpans i ve Revers ib l e I T I Wave le t T r ans f o rms They said the birds refused to sing and the thermometer fell suddenly as if God Himself had His breath stolen away. No one there dared speak aloud, as much in shame as in sorrow. They uncovered the bodies one by one. The eyes of the dead were closed as if waiting for permission to open them. Were they still dreaming of ice cream and monkey bars? Of birthday cake and no future but the afternoon? Or had their innocence been taken along with their lives buried in the cold earth so long ago? These fates seemed too cruel, even for God to allow. Or are the tragic young born again when the world's not looking? I want to believe so badly in a truth beyond our own, hidden and obscured from all but the most sensitive eyes, in the endless procession of souls, in what cannot and will not be destroyed. I want to believe we are unaware of God's eternal recompense and sadness. That we cannot see His truth. That that which is born still lives and cannot be buried in the cold earth. But only waits to be born again at God's behest... where in ancient starlight we lay in repose. —Fox Mulder, The X-Files: Closure Overview Strategies for handling the transformation of finite-extent signals in a nonexpansive manner are considered (e.g., symmetric extension, per-displace-step extension). Two families of symmetry-preserving transforms (which are com-patible with symmetric extension) are introduced and studied. We characterize the transforms belonging to these families. Some new reversible ITI structures that are useful for constructing symmetry-preserving transforms are also proposed. A simple search-based design technique is explored as means for finding low-complexity transforms in the above-mentioned families and is shown to yield good results. With this method, new effective transforms are found in addition to ones already reported in the literature. 4.1 Introduction Most frequently, wavelet transforms are defined in such a way as to operate on signals of infinite extent. In practice, however, we must almost invariably deal with finite-extent signals. Therefore, we require some means for adapting wavelet transforms to handle such signals. A reversible ITI wavelet transform is associated with a UMD filter bank of the form shown earlier in Figure 3.12. Obviously, one of the key processes involved in the computation of a such a transform is filtering. When a finite-extent signal is filtered, however, we are often faced with the so called boundary filtering problem. That is, in order to calculate the output of a filter when filtering near signal boundaries, we typically need to evaluate the signal at indices for which it is not defined. Therefore, we must alter the behavior of the transform near signal boundaries to eliminate the need for such unknown quantities. This objective can be accomplished by either changing the filters employed near the boundaries (i.e., boundary filtering) or extending the signal (i.e., signal extension). In this manner, we can eliminate the need to evaluate a signal outside of its region of definition. Chapter 4. Nonexpansive Reversible I T I Wavelet Transforms 62 Of course, signal extension and boundary filtering must be performed with care in order to avoid destroying the invertibility of a transform. Moreover, in applications such as signal coding, we almost always want to employ a transform that is nonexpansive. As a matter of terminology, a transform is said to be nonexpansive if its application to a signal comprised of N samples always yields a result that can be completely characterized by no more than N transform coefficients. In other words, a transform is nonexpansive if its use does not cause any net growth in the number of samples being processed. In this chapter, we study nonexpansive reversible ITI wavelet transforms in some detail. We begin, in Section 4.2, by discussing several signal extension techniques (e.g., periodic extension, symmetric extension, and per-displace-step extension). In so doing, we introduce the notion of a transform that preserves symmetry, and explain the significance of this property. Next, in Section 4.3, we study two families of symmetry-preserving transforms. In the context of these transform families, we also consider the relationship between symmetric and per-displace-step extension. In Section 4.4, we propose a simple technique for designing effective low-complexity transforms from the two families mentioned above. Lastly, Section 4.5 summarizes the results of this chapter. In passing, we note that some of our work presented in this chapter has also been published in [20,22,21,12,16,10]. 4.2 Extension Methods As mentioned above, one way in which to overcome the boundary filtering problem is to extend the signal being processed. In what follows, we introduce three different signal extension methods for UMD filter banks, namely periodic extension, symmetric extension, and per-displace-step extension. We also examine how the various techniques can be exploited in order to construct nonexpansive transforms. As we shall see, some of these techniques are more generally applicable than others. 4.2.1 Periodic Extension One of the simplest techniques for handling finite-extent signals is periodic extension. This technique is based on the key observation that many systems, when excited by a periodic input, always produce a periodic output (or outputs). Consider a reversible ITI wavelet transform based on the GRITIT framework. Such a transform is computed using a structure of the form shown in Figure 3.12. Let us now consider the ID case. Suppose that we are given a finite-length signal x[n] defined for « = 0,1,... , N - 1. We then choose x[n], the input to the analysis filter bank, as the following periodic extension of x[n]: x[n] = x [ m o d ( n - K , N ) ] , K & 1 (i.e., x[n] is defined such that x[n] = x[n + N]). For illustrative purposes, a simple example of periodic extension is given in Figure 4.1 for N = 4. Provided that M \ N (i.e., M divides AO, the output of each of the M downsamplers will be ^ periodic. Let us assume that none of the displace operations employed are recursive in nature. In this case, the polyphase operations preserve periodicity, and each of the M subband signals (i.e., {y,[n]}) will also be ^-periodic. Thus, we can completely characterize the subband signals by = N sample values. In this manner, we obtain a nonexpansive transform for a finite-length signal. Obviously, the above constraint on M and N may not necessarily be satisfied. That is, we may have that M )(N. For example, consider a transform based on a two-channel UMD filter bank (i.e., M — 2). In this case, the divisibility constraint is satisfied only if N is even. Clearly, in some applications, however, A' may be odd. Thus, we cannot always use periodic extension to obtain a nonexpansive transform. Furthermore, periodic extension can potentially result in an extended signal with large jumps at the splice points between periods. Such discontinuities introduce additional high-frequency content into the signal being processed, which is often undesirable in applications like signal coding. For the reasons outlined above, periodic extension is often not the most attractive method to employ. Periodic extension also has applicability in the D-dimensional case where D>2. Again, we have constraints sim-ilar to the ID case (i.e., the downsampler outputs must be periodic with no net growth in the number of nonredundant samples). 4.2.2 Symmetric Extension Another technique for handing finite-extent signals is symmetric extension. The use of symmetric extension with UMD filter banks was first proposed by Smith and Eddins [122], and subsequently received considerable attention Chapter 4. Nonexpansive Reversible I T I Wavelet Transforms 63 x,[n] 0 1 2 3 0 1 2 3 4 5 6 7 (a) (b) Figure 4.1: Periodic extension of a signal, (a) Original signal, (b) Periodic extension of signal. x[n] x[n] 0 1 2 3 4 (a) 0 1 2 3 4 5 (b) 0 1 2 |3 o 0 1 2 3 (c) (d) Figure 4.2: Types of signal symmetries, (a) WS symmetry, (b) HS symmetry, (c) WS antisymmetry, and (d) HS antisymmetry. in the literature (e.g., [123,99,34,33, I]). With this technique, a signal is extended so that it is both symmetric and periodic. Then, both the symmetry and periodicity properties are exploited. Consider a ID signal with symmetry. Such a signal can be symmetric or antisymmetric. In each case, the point of symmetry can either be coincident with a single sample or lie midway between two samples. These two types of symmetry point locations are referred to as whole sample (WS) and half sample (HS), respectively. Examples of various types of symmetries are illustrated in Figure 4.2. The symmetry centers for periodic signals always appear in pairs. If the period of a signal is even, the symmetry centers in a pair will be either both WS or both HS. If the period is odd, one will be WS and the other HS. More than one way to symmetrically extend a signal is possible. The main difference is in the number of times the first and last sample of the original signal are repeated in the extended signal. Here, we use the notation ( k , l ) to indicate that the first and last samples are repeated k and / times, respectively. For an input signal of length N, (&,/)-symmetric extension yields a (2N + k + l — 2)-periodic signal as output. Some illustrative examples of various types of symmetric extension are shown in Figure 4.3. As will be seen later, (1,1)- and (2,2)-symmetric extension are of the most interest in the context of our work herein. Consider now the UMD filter bank shown in Figure 3.12. More specifically, let us consider the two-channel case (i.e., M — 2). Suppose that we have a finite-length signal x[n] defined for n = 0,1,...,N — 1. We then choose x[n], the input to the analysis filter bank, as a symmetric extension of x[n]. Thus, the period of x[n] is (approximately) 2N. Symmetric extension is usually employed only with transforms that preserve signal symmetry. Therefore, let us assume that the filter bank preserves symmetry and periodicity. From this assumption, we have that each of the subband signals is symmetric (or antisymmetric) with a period of (approximately) N. Since each of the subband signals has symmetry, only (about) one half of the samples of a single period are required to characterize the signal. In other words, we only need (about) ^ samples from each subband signal. Thus, we need a total of J + J = N samples. Therefore, the transform is nonexpansive. Unfortunately, symmetric extension cannot always be used to obtain nonexpansive transforms. For example, this approach requires a UMD filter bank that preserves signal symmetry. This restriction is quite limiting since many Chapter 4. Nonexpansive Reversible I T I Wavelet Transforms 64 x[n] •M nJ 9 g 0 1 2 3 (a) 0 1 2 3 4 5 6 7 8 9 10 11 (b) 1 1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 (C) Figure 4.3: Symmetric extension examples, (a) Original signal. Extended signal obtained with (b) (1, l)-symmetric extension, and (c) (2,2)-symmetric extension. wavelet transforms do not have an underlying UMD filter bank that preserves symmetry. Also, symmetric extension is more difficult to apply in the ©-dimensional case where D>2. This is, in part, due to the many more types of symmetry that are possible for multidimensional signals. Symmetric extension is considered, for example, in the 2D case in [ 6 4 ] . 4.2.3 Per-Displace-Step Extension Although periodic and symmetric extension are quite popular, other simple extension schemes can also be devised. One extension technique used in conjunction with GRITIT-based (or lifting-based) transforms is the per-displace-step (PDS) extension method. This technique is mentioned briefly in [ 1 6 ] and used in the SBTLIB [ 1 0 ] and JasPer [ 1 4 ] software. More recently, this method has also been described in [32,31 ] (under the name of "iterated extension"). Consider the displace operation in the two-channel case as illustrated in Figure 4.4. With PDS extension, signal extension is performed at the input to each displace filter rather than being performed at the input to the filter bank. That is, each displace step on the analysis side of the filter bank has the form shown in Figure 4.5(a). In the diagram, po[n] and qo[n] represent intermediate lowpass channel signals, p \ [n] and q\ [n] represent intermediate highpass channel signals, and Q is a rounding operator. The parameter d determines the channel modified by the displace operation. On the synthesis side of the filter bank, the displace operation of Figure 4.5(a) has a corresponding inverse of the form shown in Figure 4.5(b). Clearly, the same extension values can be generated on both the analysis and synthesis sides (since = q\-d[n\). Thus, this scheme can generate reversible transforms. The PDS extension technique is extremely powerful, as it can always be used to form a nonexpansive transform. This method can be applied in the case of UMD filter banks with any arbitrary displace-step filters. That is, there is no need for the filter bank to preserve signal symmetry (unlike in the case of symmetric extension). Furthermore, one can easily apply PDS extension in the D-dimensional case (where D > 2). 4.3 Symmetry-Preserving Transforms As explained earlier, it is often desirable to employ transforms that preserve symmetry (i.e., transforms that yield sym-metric/antisymmetric output signals for symmetric/antisymmetric input signals). In particular, symmetry-preserving Chapter 4. Nonexpansive Reversible I T I Wavelet Transforms 65 G(z) *k[n] _ yt[n] G(z) k[n] yi-k PdW ( ^ (qd[n Q Figure 4.4: Displace step in the two-channel case. Extend <7i-(a) s Q I G(z) Pd[n\ Extend Pl-c (b) Figure 4.5: Structure of a displace step in the case of PDS extension, (a) Forward displace step and (b) inverse displace step. transforms have the advantage of being compatible with symmetric extension techniques [123,99,33,34], allowing signals of arbitrary length to be handled in a nonexpansive manner. In what follows, we introduce two families of symmetry-preserving reversible ITI wavelet transforms. Then, we proceed to study the characteristics of these families and their constituent transforms. Several interesting results are presented, which provide new insights into the transforms belonging to each family. During the course of our anal-ysis, we also develop some new reversible ITI structures that are useful for the construction of symmetry-preserving transforms. We also examine the relationship between symmetric extension and PDS extension. The presentation of this work is structured as follows. In Section 4.3.1, the two families of transforms under consideration are introduced. We proceed to show, in Section 4.3.2, that transforms from these families are compatible with symmetric extension, and can be used to handle arbitrary-length signals in a nonexpansive manner. The base filter bank associated with one of the transform families is examined in Section 4.3.3, with the filter bank's symmetry-preserving properties being of particular interest. This is followed by a more detailed study of the two transform families in Sections 4.3.4 and 4.3.5. Also, of interest, the relationship between symmetric extension and PDS extension is examined in Section 4.3.6. 4.3.1 Transform Families In our work, we consider two families of symmetry-preserving reversible ITI wavelet transforms. Both are derived from the lifting-based parameterizations of linear-phase filter banks presented in [12], and have the general form shown in Figure 4.6. In the figure, the {Qk]f^1 are rounding operators, and the analysis and synthesis polyphase filtering networks each consist of 2X lifting step filters { A t } ^ 1 . As we will demonstrate, by choosing the lifting step filters { A * } ^ 1 wisely, one can construct filter banks that not only yield symmetric/antisymmetric subband signals (i.e., yo[n] and yi[n}) for an appropriately chosen input signal (i.e., x[n]), but also yield intermediate signals (i.e., {"tMijfcio1' { v *M } i tM ) 2 ) t n a t 3 1 6 a ^ o r m o s t ' y symmetric/antisymmetric. Chapter 4. Nonexpansive Reversible I T I Wavelet Transforms 66 12 "J | "'J"! (+) "»2^ . . . M2L-2t"l ^ "a-jiW Q yofc 42 Ga-i Aa-i(z) yi["J G21-2 •<i> vo[«J vi[n] v2[nj v a - 2 l " J (a) yofo] Q »2x-)M "2^-2M ,. "2[y] Q mjn; Ao(z) 1 T Co 5 V2X-2M v2[n] v,[n) ^ v0[n] (b) Ao{z) Qi Q2X-1 I Qo Gu-2 ^2X-l(z) T2 12 J i ["] Figure 4.6: Lifting realization of a reversible ITI wavelet transform, (a) Forward and (b) inverse transforms. 4.3.1.1 OLASF Family The first family of reversible LTI wavelet transforms is associated with a linear-phase filter bank having odd-length analysis/synthesis niters, and has been described in [12]. For convenience, we will refer to this family as the odd-length analysis/synthesis filter (OLASF) parameterization. In the case of this family, the parameters in Figure 4.6 are constrained as follows: The filters { A t } ^ 1 3 1 6 chosen to have transfer functions of the form , , x _ /x! i 0 " 2 ) / 2 Mz - ' '+0 for e v e n * ^'WS^a^-'+z1) forodd* i 2 X - l (4.1) where the {£*}*=ol a r e a 1 1 e v e n i n t e § e r s > a n d t h e rounding operators {Qk}^1 are chosen arbitrarily (e.g., as the floor, biased floor, ceiling, biased ceiling, truncation, biased truncation, or RAFZ function). Without loss of generality, we assume that none of the {Akiz)}^2 are identically zero. That is, only A0(z) and A2x-i(z) may be identically zero. Some well known members of this family include the 5/3, 9/7-M, 5/11-C, 13/7-T, 13/7-C, and 9/7-F transforms discussed in Chapter 5. 4.3.1.2 ELASF Family The second family of reversible ITI wavelet transforms is associated with a linear-phase filter bank having even-length analysis/synthesis filters, and has been described in [12]. For convenience, we will refer to this family as the even-length analysis/synthesis filter (ELASF) parameterization. In the case of this family, the parameters in Figure 4.6 are constrained as follows: The filters { A * } ^ 1 are chosen to have transfer functions of the form -1 X2+M(z) for k = 0 for it = 1 for k > 2 (4.2a) where 4to = 2&~1)/2M*-'-*•'), (4.2b) Chapter 4. Nonexpansive Reversible I T I Wavelet Transforms 67 ffcjL 12 12 (a) 1 ei yo["J A 2 (z) G2 JEfejL 12 12 Q i -1 (b) 1 " A 2 (z) > 1 G2 Figure 4.7: Two different reversible ITI approximations of the 6/14 transform. The analysis filter bank based on the (a) commonly referenced parameterization and (b) our more general parameterization. Ai(z) = — fg(z — z - 1 )» A2(z) = -rg ((z2 — z - 2 ) — 6(z — z _ 1 ) ) , Q\ is integer-bias invariant, Q\ is arbitrary, and Q 2 is odd. A > 1, and the { L k } f ^ } are all odd integers; and the rounding operators {Qk} are chosen such that Qo is the identity, Qi is integer-bias invariant (e.g., the floor, biased floor, ceiling, or biased ceiling function), and the remaining rounding operators { Q k } 2 ^ . 1 are odd (e.g., the truncation, biased truncation, or RAFZ function) for even k and arbitrary for odd k. Without loss of generality, we assume that none of the {-^(z)} 2^ 2 are identically zero. That is, only A \ ( z ) and A 2 \ _ i (z) may be identically zero. Some well known members of this family include the 2/6, 2/10, and 6/14 transforms discussed in Chapter 5, in addition to the S transform. The above transform family is a generalization of ones commonly considered in the literature (e.g., [32]). Our proposed family is a generalization in two respects. First, in the case of transform families with a similar structure, it is often suggested (e.g., as in [32]) that the { Q k } f ^ i must all be chosen as odd functions. Such a choice, however, is overly restrictive. The rounding operators {G*} 2^' o n t y n e e d t 0 ^ e c n o s e n as odd functions for even k. Second, and more fundamentally, the related families commonly considered in the literature (e.g., as in [32]) are essentially a special case of our proposed family with A \ (z) = 0. The additional flexibility afforded by the introduction of the A\ (z) parameter has important practical implications. In many cases, by using our more general parameterization in (4.2), the number of lifting steps can be reduced by one. By reducing the number of lifting steps, the number of rounding operations is also decreased, leading to a reduction in rounding error. Consider, for example, the two (ITI) filter banks shown in Figure 4.7. One can easily verify that both filter banks approximate the same linear transform, namely the one associated with the 6/14 transform in [16] which is known to be effective for image coding purposes. Both filter banks are reversible, as both employ lifting/ladder networks for polyphase filtering, but due to the nonlinearities introduced by rounding, the two filter banks are not equivalent. Comparing the filter banks of Figures 4.7(a) and 4.7(b), we can see that the former has one more rounding operation than the latter. For this reason, the filter bank in Figure 4.7(b) will normally have better approximation properties (i.e., smaller rounding error) than the filter bank of Figure 4.7(a). Although not immediately obvious, like the filter bank in Figure 4.7(a), the filter bank in Figure 4.7(b) also preserves symmetry, as we will show in a later section. 4.3.2 Transforms and Symmetric Extension In the previous section, we introduced the OLASF and ELASF families of reversible ITI wavelet transforms. Now, we show that the transforms from these two families can be used with symmetric extension in order to handle signals of arbitrary length in a nonexpansive manner. 4.3.2.1 O L A S F Case Consider a filter bank of the form shown in Figure 4.6 that is constrained to be of the OLASF type as defined by (4.1). Suppose that we are given a signal x[n] defined for n = 0,1,..., N — 1. By using symmetric extension, one can form a nonexpansive transform, as demonstrated by the proposition and accompanying proof below. Proposition 6 (Compatibility of transforms from the O L A S F family with symmetric extension). Consider a filter bank of the form shown in Figure 4.6 that is constrained to be of the OLASF type as defined by (4.1). Suppose now Chapter 4. Nonexpansive Reversible I T I Wavelet Transforms 68 that we are given a signal x[n] defined for n = 0,1,..., /V — 1. I f we choose x[n), the input to the analysis filter bank, as the symmetric extension ofx[n] given by x[n] =x[min(mod(n - K,2N - 2),2N - 2 - mod(n - K,2N - 2))], K G Z (i.e., x[n] is defined such that x[n] = x[n + 2N — 2] andx[n] — x[2K — n]), then 1. yo[n] is completely characterized by its No samples at indices \ j ~ \ , [§] +1,..., [^+^f~1 j ; 2. yi [n] is completely characterized by its N\ samples at indices , [^ f^ l + !>•••> [^y-^J ; and 3. No + N\ = N (i.e., the transform is nonexpansive). Proof. Consider the OLASF-type system associated with Figure 4.6. Evidently, x[n] is (2N — 2)-periodic and symmet-ric about K and K + N— 1 (with additional pairs of symmetry points following from periodicity). From the properties of downsampling, one can show that uo[n] is (N— 1)-periodic and symmetric about § and K + I ^ ~ ' , and vo[n] is (N — 1)-periodic and symmetric about and K + I ^ ~ 2 . Consider now the first lifting step (associated with filter Ao). Since, from (4.1), the filter Ao has a group delay of — 5 and the rounding operator Qo preserves symmetry, the adder (follow-ing Go) is summing two (N — l)-periodic symmetric signals with the same symmetry centers. Thus, the adder output, vi[n], is also (N — l)-periodic and symmetric with the same symmetry centers, namely and K+2 ~2 • Consider now the second lifting step (associated with filter Ai). Since, from (4.1), the filter A\ has a group delay of \ and the rounding operator Gi preserves symmetry, both adder inputs are (N — l)-periodic symmetric signals with the same symmetry centers. Therefore, the adder output, U2[n], must also be (N — l)-periodic and symmetric with the same symmetry centers, namely * and K + ^ ~ x . By repeating the above reasoning for each of the remaining lifting steps, we have that the {Mn]}^1 and yo[w] are all ( N - l)-periodic and symmetric about \ and K+2~l• Likewise, the {v*M}fM)2 andyi[«] are all (A'— l)-periodic and symmetric about and K+2"2• By examining the form of the various signals, we can see that the {wtMj^o 1 a n c ^ ^ completely character-ized by their samples at indices \j], [§ ] + 1,..., j , for a total of No independent samples, where "0=1^]-[§1 + 1. ( 4.3) Similarly, we have that the {v* [n] }2^Q2 and y\[n] are completely characterized by their samples at indices [^ "1, \^f^~\ + I,..., [ K + 2 ~ 2 \ ' f ° r a t o t a l of Ni independent samples, where Using (2.11) and (2.10), we can further simplify (4.3) and (4.4) to obtain N 0 = < j for even N ^ i l for odd N , even K and N\ ^ f i for odd A', odd K , [ \ for even N ^ for odd AT, even K (4.5) for odd N , odd K . Thus, from (4.5), we have that, regardless of the parities of K and A', No + N \ = N . In other words, the total number of independent samples in the transformed signal is always equal to the number of samples in the original signal, and the resulting transform is, therefore, nonexpansive. • 4.3.2.2 ELASF Case Consider a filter bank of the form shown in Figure 4.6 that is constrained to be of the ELASF type as defined by (4.2). Suppose that we are given a signal x[n] defined for n = 0,1,... ,N — 1. By using symmetric extension, one can form a nonexpansive transform, as demonstrated by the proposition and accompanying proof below. Proposition 7 (Compatibility of transforms from the ELASF family with symmetric extension). C o n s i d e r a filter bank of the form shown in Figure 4.6 that is constrained to be of the ELASF type as defined by (4.2). Suppose now Chapter 4. Nonexpansive Reversible I T I Wavelet Transforms 69 that we are given a signal x[n] defined f o r n = 0,l,...,N—l. I f we choose x[n], the input to the analysis filter bank, as the symmetric extension ofx[n] given by x [ n ] = x [ m m ( m o d ( n - K , 2 N ) , 2 N - l - m o d ( n - K , 2 N ) ) } , KeZ (4.6) (i.e., x[n] is defined such thatx[n] = Jt[n + 2A'] andx[n] =x[2K— 1 — n]), then 1. yo[n] is completely characterized by its No samples at indices , + !>•••) [/ir+2V~1J>' 2. yi [n] is completely characterized by its Ni samples at indices [ j] , fy] + 1,..., [K+%~2J»' a n d 3. No + Ni = N (i.e., the transform is nonexpansive). Proof. Consider the ELASF-type system associated with Figure 4.6. For the time being, we assume that W2M is iV-periodic and symmetric about and K + 2 ~ 1 , and v2[n] is A'-periodic and antisymmetric with the same symmetry points as u 2 [ n ] . These assumptions will be proven valid in Section 4.3.3. Consider now the third lifting step (associated with filter A 2 ) . Since, from (4.2), the filter A 2 has a group delay of zero, and the rounding operator Q 2 preserves antisymmetry (since it is odd, by assumption), the adder is summing two A'-periodic antisymmetric signals with the same symmetry centers. Thus, the adder output is also A'-periodic and antisymmetric with the same symmetry centers. Consider the fourth lifting step (associated with filter A3). Since, from (4.2), the filter A3 has a group delay of zero, and the rounding operator Q3 preserves symmetry, the adder is summing two A'-periodic symmetric signals with the same symmetry centers. Thus, the adder output is also A'-periodic and symmetric with the same symmetry centers. It is important to note that, for odd k, Qk need not be odd, contrary to the suggestions of some, in the case of similar parameterizations. By repeating the above reasoning for each of the remaining lifting steps, we have that the {^[n]}^ 1 a n u yo[n] are A'-periodic and symmetric with symmetry centers and K + 2 ~ l , and the {v*[n]}?^~2 and yi \n] are A'-periodic and antisymmetric with the same symmetry centers. By examining the form of the various signals, we note that the {^[n]} 2^' andyo[«] are completely characterized by their samples at indices [^ j—1 > + x> • • • > [ / r + 2~ ' J > f ° r a t o t a l of No independent samples, where N 0 = [ ^ \ - \ ^ ] + l . (4.7) Similarly, we have that the {v^n]} 2 ^ 2 and y\[n] are completely characterized by their samples at indices [^], + 1,..., [ K + 2 ~ 2 j , for a total of N\ independent samples, where A ^ i ^ J - f f l + 1. (4.8) Using (2.10) and (2.11), we can further simplify (4.7) and (4.8) to obtain for A' odd for N even, K even and N\ — { ^ for N even, K odd, ( N-1 2 2 for A' odd for N even, K even (4.9) ^ for N odd, K odd. Thus, from (4.9), we have that, regardless of the parities of K and A', N 0 + N i = N . In other words, the total number of independent samples in the transformed signal is always equal to the number of samples in the original signal, and the resulting transform is, therefore, nonexpansive. • 4.3.3 Symmetry Preservation in the ELASF Base Filter Bank Consider a filter bank of the form shown in Figure 4.6 that is constrained to be of the ELASF type as defined by (4.2) with X = 1. The analysis side of such a filter bank has the form illustrated in Figure 4.8. In the previous section, we assumed that if x[n] is generated using (4.6), then yo[n] is A'-periodic and symmetric about and K + N ~ x , and y\ [n] is A'-periodic and antisymmetric with the same symmetry centers as yo[n]. We will now prove these assumptions to be valid. Chapter 4. Nonexpansive Reversible I T I Wavelet Transforms 70 12 Knjn] »ijn] Q (yo[n -1 12 ^H-A.(z) s vifn] y i H0(z) 12 Figure 4.8: Base analysis filter bank for the ELASF family of transforms. .yi[«] yiM Hdz) 12 T2 Fdz) (a) (b) Figure 4.9: The general structure of a two-channel UMD filter bank, (a) Analysis filter bank and (b) synthesis filter bank. 4.3.3.1 Symmetry Preservation We begin by considering a slightly different filter bank of the form shown in Figure 4.10. This filter bank can be represented in canonical form as shown in Figure 4.9. Let us denote the transfer functions of the lowpass and highpass analysis filters associated with the canonical form of the filter bank as Ho{z) and H\ (z), respectively. Using the noble identities, we can show that //o(z)=2(z+l) + (z-l)Ai(z 2 ) and //,(z) = z - l . Due to the form of A\ (z) in (4.2), it follows that the filter Ho has an impulse response symmetric about — \ . Clearly, the filter H\ has an impulse response that is antisymmetric about — \ . Consequently, if x[n\ is generated using (4.6), we have that yo[n] is A'-periodic and symmetric about and y\ [n] is A'-periodic and antisymmetric with the same symmetry center as yo[«]- Since yo\n] and yi[n] are A'-periodic, it follows that both signals must also have another symmetry center at K+2~l • Now, suppose that we round the lowpass subband output, so as to obtain the system shown in Figure 4.11. For any (non-pathological) rounding function, this process will maintain signal symmetry. Moreover, if the rounding operator, Q, is integer-bias invariant, we can equivalently move the rounding operator to the input side of the adder, resulting in the system in Figure 4.8. Consequently, in the case of the system in Figure 4.8, if the input x[n] has the form of (4.6), yo[n] must be A'-periodic and symmetric about and y + 2 ~ ' . (Since yi [n] has not been changed by the introduction of the rounding operator Q, this signal's symmetry properties are obviously unchanged from above.) Thus, we have shown our above assumptions about the form of yo[n] and yi [n] to be correct. 4.3.3.2 Modifying the E L A S F Base Filter Bank As demonstrated above, the system of Figure 4.11 has the symmetry-preserving properties that we desire, regardless of whether or not the rounding operator Q is integer-bias invariant. If Q is not integer-bias invariant, however, it is not clear that the resulting transform will necessarily have an inverse. In what follows, we will examine this question of transform invertibility more closely. Consider a network of the form shown in Figure 4.12, where the co[n], c\[n], do[n], and d\ [n] are integer sequences and Q is a rounding operator. Obviously, the network in Figure 4.11 can be inverted if the network in Figure 4.12 can be inverted. In turn, the latter network can be inverted if we can invert the network in Figure 4.13, where x,y € Z and a £ i . Thus, we have reduced the overall invertibility problem to a much simpler one. That is, the network Chapter 4. Nonexpansive Reversible ITI Wavelet Transforms 71 12 42 •<t> z -1 42 42 yi[«] Figure 4.10: Linear version of the base analysis filter bank for the ELASF family of transforms. c0[n G(z) ci[n] do[n] Figure 4.12: Network consisting of a ladder step and rounding unit. Figure 4.11: Modified base analysis filter bank for the ELASF family of transforms. Figure 4.13: Network consisting of an adder and rounding unit. in Figure 4.11 is invertible if we can invert (i.e., solve uniquely for x in terms of y) the equation y = Q{a+x) (4.10) where x,y £ Z and a is a real constant. We assert that this equation has an inverse if Q is chosen as either the biased truncation or RAFZ function, but not if Q is chosen as the truncation function. This assertion is proven by the three propositions below. Proposition 8. Consider the equation y = trunc(0C + x) (4.11) where x,y £ Z and a is a real constant. (This equation is a special case of (4.10) where Q is chosen as the truncation function.) If a £ Z, the above equation is not invertible (i.e., one cannot always uniquely solve for x in terms ofy). Proof. Since a ^ Z, from the definition of the floor function, we have 0 < a — [aj < 1 which implies - 1 < a - | a j - 1 < 0. The preceding two relationships imply, respectively, that trunc(oc— |_aj) = 0 and trunc(a - |_aj — 1) = 0. Thus, for any a ^ Z, we have that trunc(a+x) — 0 for both x = - [aj and x — — \ a\ — 1. Since two distinct values for x both yield the same value for y, (4.11) cannot be invertible. In passing, we note that, obviously, (4.11) is invertible if a 6 Z. In this case, (4.11) is simply equivalent to y = a+x, which is trivially invertible. • Proposition 9. Consider the equation y = btrunc(a+:c) (4.12) where x,y £ Z and a is a real constant. (This equation is a special case of (4. JO) where Q is chosen as the biased truncation function.) For any choice of a, the above equation is invertible (i.e., one can always uniquely solve for x in terms ofy). Moreover, the inverse is given by | y — a — j sgny if a is an odd integer multiple of 5 ^ \y — bfloora otherwise. Chapter 4. Nonexpansive Reversible ITI Wavelet Transforms 72 Proof. Let us begin by considering the case that a is an odd integer multiple of \ . From the definition of the btrunc function in (2.7), we can rewrite (4.12) as y = btrunc (a + x) f[a+*+il fora+x>0 ( 4 1 4 ) 1 [a + jc-^] for a+x < 0. Since a is an odd integer multiple of we know that cx + x+j and a- l -*- 3 are both integers. Thus, we can simplify (4.14) as follows: _ja+x+2- fora+x>0 l a + j c - j fora+*<0 = a+x+ \sgn(a+x). From (4.12), we can deduce that sgn(a + ;t) = sgny, since sgnbtrunc(oc + *) = sgn(a-f-x) if \a + x\ > \ which must be the case here. Consequently, we can solve for x in terms of y in the above equation to obtain x — y — a—j sgny. Thus, we have proven (4.13) to be correct in the case that a is an odd integer multiple of 5 -l 2 Now let us consider the case that a is not an odd integer multiple of A. Using (2.7) and Proposition 3 (in Sec-tion 3.2.2), we can simplify (4.12) as follows: y = btrunc (a + x) = bfloor(a + ^ ) = x + bfloora. Solving for x in terms of y in the preceding equation, we obtain x = y — bfloora. Thus, we have proven (4.13) to be correct in the case that a is not an odd integer multiple of j , completing the proof. • Proposition 10. Consider the equation y = rafz(a-f-*) (4.15) where x,y £ Z and a is a real constant. (This equation is a special case of (4.10) where Q is chosen as the RAFZ function.) For any choice of a, the above equation is invertible (i.e., one can always uniquely solve for x in terms of y). Moreover, the inverse is given by f b - a j fory>0 \\y-a] fory<0. Proof. Using the definition of the RAFZ function in (2.8), we can rewrite (4.15) as follows: y =rafz(x + a) _j\x + d] forx+a>0 1|*-r-aJ for;c-|-a<0 fjc+fa] for;c-l-a>0 jjc+LaJ for;c+a<0. Solving for x in the above equation, we obtain l y - f a l forx + a>0 ( 4 1 ? ) l y - | a j for;c-|-a<0. Chapter 4. Nonexpansive Reversible I T I Wavelet Transforms 73 12 1 w 12 12 12 0> Win] Figure 4.14: Linear version of the base analysis filter bank Figure 4.15: Modified base analysis filter bank for the for the new ELASF-like family of transforms. ELASF-like family of transforms. From the definition of the RAFZ function, however, we know that sgn(rafza) — sgna for all a 6 R (i.e., the RAFZ function preserves signedness). Consequently, we have that sgn(*-foc) = sgny. Using this fact and algebraic manipu-lation, we can rewrite (4.17) as follows: fory >0 fory <0 fory >0 for y < 0. Thus, we have solved for x in terms of y and a, proving the correctness of (4.16). • From the above propositions, we consequently have that the the network in Figure 4.11 can be inverted if Q is chosen as the biased truncation or RAFZ function, but not if Q is chosen as the truncation function. We can exploit this knowledge in order to gain increased flexibility in the construction of symmetry-preserving reversible ITI transforms. For example, consider the linear-phase analysis filter bank shown in Figure 4.14, where A i (z) is of the form given in (4.2b). Due to the linear-phase property of the analysis filters, this linear filter bank is symmetry preserving. In particular, one can show that if x[n] is generated using (4.6), then yo[«] is symmetric and yi [n] is antisymmetric. We can, therefore, obtain a symmetry-preserving ITI version of this filter bank by simply rounding the output of the highpass subband signal (to an integer value). In so doing, we obtain the new structure shown in Figure 4.15, where Q is a rounding operator. Since we would like to preserve antisymmetry in the highpass subband output, we must choose Q to be odd. This has the implication, however, that Q cannot be integer-bias invariant. Consequently, unlike in the case of the ELASF base filter bank of Figure 4.11 (considered earlier in this section), we cannot simply move the rounding unit before the adder. We know from our previous results, however, that if Q is chosen as either the biased truncation or RAFZ function, the resulting filter bank (in Figure 4.15) must be reversible. Furthermore, in each case, the corresponding inverse filter bank can be easily deduced by using the results of Propositions 9 and 10 above. Thus, we have, in effect, constructed a new symmetry-preserving base filter bank, similar to the one used in the ELASF family. Moreover, the newly proposed reversible structures may be useful for the construction of more general symmetry-preserving reversible ITI transforms. Summary In summary, this section first examined the symmetry-preservation properties of the ELASF base filter bank. This, in turn, led us to consider the properties of a slightly different base filter bank. By studying this new filter bank, we were able to propose some new reversible ITI structures that are useful for constructing symmetry-preserving transforms. 4.3.4 O L A S F Family One can show that any PR (i.e., perfect-reconstruction) linear-phase FTR ID two-channel filter bank with odd-length analysis/synthesis filters has a corresponding OLASF parameterization, provided that the analysis and synthesis filters are normalized appropriately (via a scaling and shift of their impulse responses). For example, the lowpass and high-pass analysis filters must be normalized such that they have impulse responses centered about 0 and —1, respectively. The proof is constructive by a matrix Euclidean algorithm (e.g., by using an algorithm based on that described in [49]). Chapter 4. Nonexpansive Reversible I T I Wavelet Transforms 74 4.3.5 E L A S F Family Only a subset of all PR linear-phase FIR ID two-channel filter banks with even-length analysis/synthesis filters have corresponding ELASF parameterizations. Moreover, this subset is quite small, as there are many such filter banks that cannot be realized in this way. In what follows, we examine some of the characteristics of transforms in the ELASF family, and in so doing, demonstrate some of the reasons for the incompleteness of this parameterization. 4.3.5.1 Transform Properties Consider the linear version of a filter bank of the form shown in Figure 4.6(a) that is constrained to be of the ELASF type as defined by (4.2). (By "linear version", we mean the filter bank obtained by simply setting all of the {fo} 2 ^ 1 equal to the identity.) This filter bank can be represented in canonical form as shown in Figure 4.9. Let us denote the corresponding lowpass and highpass analysis filter transfer functions as HQ(Z) and Hi (z), respectively. Further, we define the transfer matrix for part of the polyphase filtering network as follows: (4.18) n 1 A2t+i(z) 0 1 Alk(z) l j / [0 1 1) (4.19a) (4.19b) Using noble identities and straightforward algebraic manipulation, we can show K)]=p^[0t]^ m=^ )[^ For convenience, can rewrite the preceding equation as follows: H0(z) =$(z + l)P0,o(z2) + (z- l)P0,i(z2) and #i(z) =(z- l)/>i,i(z2) + £(z + l)Pi,o(z2). A detailed analysis of (4.18) reveals that P(z) has the following important properties: 1- o^,o(z) and Pi,i (z) both have coefficient sequences that are symmetric about 0; 2. Po,i (z) and Pi,o(z) both have coefficient sequences that are antisymmetric about 0; 3. For i = 0,1 and j = 0,1, degP;j (z) is even, except when P;,;(z) = 0; deg/fyoCz) > 0 and deg/^i (z) > 0. 4. If at least one of the {Ai(z)} is not identically zero, we have that either: degPo,o(z) >degPrj,i(z) and deg P\ $(£) > degPM(z); or degP0fl(z) < degP0,i(z) and degPi,o(z) < degPM(z); 5. degP0,o(z) j= degP0,i(z), degPi,0(z) ^ degPi,i(z), degP0,o(z) ^ degPi,o(z). anddegP0,i(z) degA,i(z); 6. fl,i0(l) = landP M ( l) = l . The above properties can be easily deduced from Propositions 11 and 12 given below. Proposition 11. Suppose that we have a product of the form pW(z) = where N > 0, Mz) = i Mz) 0 1 1 0 Mz) i UUMz) for even i for odd i, (4.20) Chapter 4. Nonexpansive Reversible I T I Wavelet Transforms 75 Ai(z) ^ 0 , and Z ~ x A i ( z ) is antisymmetric about 0, f o r i = 0,1,... , N - 1. Then, pW(z) must be such that Z~1 P$ (z) and Z~1P^ (z) are symmetric about 0; ZT1 P$ (z) and Z~xpf^ (z) are antisymmetric about 0; f o r N ^ 0: d e g / ^ z ) < d e g ^ ( z ) , d e g ^ ( z ) < degP^z) ; (AO/ >W/ f o r even N , N ^ 0 : d e g P $ ( z ) < degP$(z), degP^z ) < degP^z ) ; •(AO, 3(A0/ >(A0/ /brodtfJV: degPg^z) >degP,7(z), degP^z) > degP^z ) ; .(AO/ 3(A0/ (4.21a) (4.21b) (4.21c) (4.2 Id) (4.21e) <o )(i) = i^ i7(1) = 1- (4.22) Proc/ The first part of the proof is by induction on N . First, consider the cases of N = 0,1,2. For these three cases, respectively, (z) is given by r i .0 U > 1 A0(z) 0 1 , and 1 A 0(z) A,(z) l+A0(zMi(z) (4.23) In the first two of the above three cases, we can easily confirm that (4.21) is satisfied. In the third case, (4.21) can also be shown to hold by using reasoning similar to that which follows. In the interest of brevity, however, we will not explicitly demonstrate this here. To complete the inductive process, we must now show that if (4.21) holds f o r N = K then (4.21) also holds for N = K+ 1. Thus, we begin by assuming that (4.21) is satisfied for N = K. From (4.20), we can write P(*+I>(z) = L^ + , )W^ + , )WJ ,(*+!)/ = < 1 AK(z) 0 1 1 0 AK(z) 1 = < p ( * > M P » ^ ( Z ) P W ( Z ) for even AT for odd 4 ?M for even K for odd K. Thus, the expression for P^K+l\z) can have one of two forms depending on the parity (i.e., evenness/oddness) of K. Consider (4.21a). First, suppose that K is even. Since PQKQ\Z) and AK(Z)P[KQ(Z) both have a coefficient sequence that is symmetric about 0, their sum PQK0+1\Z) must also have a coefficient sequence with the same symmetry. Thus, by observing P[KI+1\Z) = Pffliz), we have proven that (4.21a) holds for N = K + 1 when K is even. In a similar manner, we can also show that (4.21a) holds for N = K+ 1 when K is odd. Consider (4.21b). First, suppose that K is even. Since PQWZ) and Ajt(z)P,7(z) both have a coefficient sequence that is antisymmetric about 0, their sum PQ\+1\Z) must also have a coefficient sequence with the same symmetry. Thus, by noting P\0 '(z) = P i 0 ' (z), we have proven that (4.2 lb) holds for N = K + 1 when K is even. Using similar reasoning, we can also show that (4.2 lb) holds for N = K + 1 when K is odd. Consider now (4.21c) to (4.2le) when K is even. Since P $ ( z ) and AK(Z)P^Q (z) both have a coefficient sequence centered about 0 and deg PQK0\Z) < deg P[K0\Z), the terms with the lowest and highest powers of z in PQ* + 1 ) (Z) are contributed exclusively by AK{Z)P[KQ(Z). Consequently, we have degP$ + 1 )(z) = degAx(z)P,7(z) = deg^(z)+ deg/>^(z) . 3(ic+i), (4.24) Since PQ^(Z) and AK(z)P\y(z) both have a coefficient sequence centered about 0 and degP^Cz) < degP^/(z), the terms with the lowest and highest powers of z in PQ* + 1 ' (Z) are contributed exclusively by AK(z)P\y (z). Consequently, (*)/ Chapter 4. Nonexpansive Reversible ITI Wavelet Transforms 76 we have degpg+%) = degAK&P^iz) = degAK(z) + de g pg + 1 ) ( z ) . (4.25) From (4.24) and (4.25), we have deg^J + 1 ) ( z ) -degP 0 (* + 1 )(z) = degPf J + 1 ) ( z ) -degP, (* + 1 )(z). This, however, im-plies that (4.21c) holds for A' = K + 1 when K is even. From (4.24), since degA*(z) > 0, we have degPQKQ+l\z) > deg/"j Q+ 1^(Z), proving that the first part of (4.2 le) holds for N = K+1 when K is even. From (4.25), since degA^(z) > 0, we have degP^l\z) > degpj*+1^(z), proving that the second part of (4.2 le) holds for N = K+l when K is even. (K) (K) Consider now (4.21c) to (4.2le) when K is odd. Since P{0(z) and Ak(Z)PQ0 (z) both have a coefficient sequence centered about 0 and degPoo'(z) > degPJo^(z), the terms with the lowest and highest powers of z in P[K0+1\Z) are (K) 1 contributed exclusively by A# (z)Pg0 (Z). Consequently, we have deg/f 0 + 1 )(z) = degAK(z)P^\z) = degAK(z) + degP«+l)(z). (4.26) (K) (K) (K) (K) Since P{ ,(z) and AK(Z)PQ l (z) both have a coefficient sequence centered about 0 and degPp i (z) > degPj, (z), the (K+l) (K) terms with the lowest and highest powers of zin P{ j (z) are contributed exclusively by AK(Z)PQ ( (z). Consequently, we have degP , ( * + n (z) = degAK(z)P$\z) = degAK(z) + de g pg + 1 ) ( z ) . (4.27) From (4.26) and (4.27), we have degP, (£ + 1 )(z) - degP,(*+1)(z) = degP^(z) - degP^(z). This, however, implies that (4.21c) holds for N = K + 1 when K is odd. From (4.26), since degA*:(z) > 0, we have deg/ >^ + 1 )(z) < degp(^+ 1 ) (z), proving that the first part of (4.21d) holds for N = K +1 when K is odd. From (4.27), since degAK{z) > 0, we have degP^ + 1 ) (z) < degPJ*+1)(z), proving that the second part of (4.21d) holds for N = K + 1 when K is odd. From above, we have shown that if (4.21) holds for N = K, then it also holds for N = K+l. This completes the inductive part of the proof. Since, for i = 0,1,... ,N - 1, 2 _ 1 A,(z) is antisymmetric, A,(l) = 0 and = J. Thus, PW(1) =IN = I, and this implies that P{$(1) = 1 and P{^(1) = 1. Thus, we have that (4.22) holds for any A'. In passing, we note that conditions (4.21a) and (4.21b) together imply that for i = 0,1, j = 0,1, degP-j\z) must be even, unless pffl (z) = 0. • Proposition 12. Suppose that we have a product of the form pW(z) = where N > 0, U&Mz) (4-28) Mz) 1 o Ai(z) 1 \Ai(z) 0 1 for even i for odd i, A,(z) ^ 0, and ZJ_1A,(z) is antisymmetric about 0,for i = 0,1,... ,N — 1. Then, P("'(z) must be such that Z~1 P$ (z) and ZT1 P , ^ (z) are symmetric about 0; (4.29a) Z T 1 P ^ (z) and Z~1P^ (z) are antisymmetric about 0; (4.29b) forN^ 0: degP^(z) > degP™ (z), degP[N0\z) > degpfj(z); (4.29c) for even N, N ^ 0: degP$ (z) > degpf j (z), d e g P ^ z ) > degpf j (z); (4.29d) for odd N: degP${z) < degP[NQ\z), degpg^z) < degPJJ (z); am/ (4.29e) ^5(l) = l , p fy ( l ) = l . (4.29f) Chapter 4. Nonexpansive Reversible I T I Wavelet Transforms 77 Proof. We begin by rewriting (4.28) as where pw(z) = j(u!^t1Mz))j (4.30) Mz) i Mz) o 1 1 0 Mz) i for even i for odd i. From (4.30), we can write where ^ [ Z ) ~ KoM A.i(z)J _11'=° , W P(z) = J P ( w ) ( z ) J . From (4.32), we have that P(z) and pM(z) are related as follows: fS$(z) = Pl,i(z), pff(z) = P,,o(z), (4.31) (4.32) (4.33) P['$(z)=PbAz), and ^ (z )=f t ,o (z ) . We now observe that the product in (4.31) is of the same form considered in Proposition 11. Consequently, we can easily show, by using Proposition 11 and (4.33), that (4.29) holds. • In what follows, the above properties will be used to study various characteristics of the analysis filters such as the form of their transfer functions, their lengths, and their DC and Nyquist gains. These characteristics will also be used to explain the reasons for the incompleteness of the ELASF parameterization. 4.3.5.2 Lowpass Analysis Filter Transfer Function For convenience, let us denote the first and second terms in the expression for Ho(z) in (4.19a) as B(z) and C(z), respectively. That is, we define B{z) = \ ( z + \ ) P 0 , o ( z 2 ) and C(z) = (z- l )P 0 , i (z 2 ) . Due to the form of P0,o(z2) (which is implied by property 1 of P(z)), B(z) has a coefficient sequence b[n] that is symmetric about — \ with adjacent pairs of samples being equal in value (i.e., b[2n] = b[2n — 1]). Likewise, due to the form of Po,i(z2) (which is implied by property 2 of P ( z ) ) , C(z) has a coefficient sequence c[n] that is symmetric about — \ with adjacent pairs of samples being equal in magnitude but opposite in sign (i.e., c[2n] = —c[2n — 1]). Suppose that Po,i (z) ^ 0. In this case, from properties 3 and 5 of P ( z ) , we know that degPo,o(z2) and deg/fyi (z2) must differ by a nonzero integer multiple of 4. Since Ho(z) = B(z) + C(z), Ho(z) must have a coefficient sequence that is symmetric about —\ and begins and ends with pairs of coefficients that are either equal or equal in magnitude but opposite in sign. In the degenerate case, in which Po,i (z) = 0, we simply have HQ(Z) = \ { z + 1). By considering the degrees of B(z) and C(z), we can also see that degflb(z) = 1 + 2max(degP0,o(z),degP0,i(z)). (4.34) Since, by property 3 of P(z), degPo,o(z) is always even and degPo.i (z) is even (except when Po,i (z) = 0), we have that deg#o(z) is odd. Thus, as had been suggested earlier, Ho is an even-length filter. 4.3.5.3 Highpass Analysis Filter Transfer Function For convenience, let us denote the first and second terms in the expression for H i ( z ) in (4.19b) as B(z) and C(z), respectively. That is, we define B(z) = ( z - l)Pi,i (z2) and C(z) = {-(z+ l)Pi, 0(z 2). Due to the form of Pi,i (z2) (which is implied by property 1 of P(z)), B(z) has a coefficient sequence b[n] that is antisymmetric about — j with adjacent pairs of samples being equal in magnitude but opposite in sign (i.e., b[2n] = —b[2n — 1]). Likewise, due to the form of Pi,o(z2) (which is implied by property 2 of P ( z ) ) , C(z) has a coefficient sequence c[n] that is antisymmetric about — 5 with adjacent pairs of samples being equal (i.e., c[2n] — c[2n — 1]). Suppose that Pi,o(z) ^ 0. In this case, from properties 3 and 5 of P(z), we know that degPi^z 2) and degPi)o(z2) must differ by a nonzero integer multiple of 4. Since H\(z) = B(z) + C ( z ) , #i(z) must have a coefficient sequence that is antisymmetric about — ^ and begins and Chapter 4. Nonexpansive Reversible I T I Wavelet Transforms 78 ends with pairs of coefficients that are either equal or equal in magnitude but opposite in sign. In the degenerate case, in which P\ ,o (z) = 0, we simply have H\ (z) = z — 1. By examining the degrees of B(z) and C(z), we can see that deg#i(z) = l+2max(deg/'i,1(z),degPlio(z)). (4.35) Since, by property 3 of P(z), degPij (z) is always even and degPi,o(z) is even (except when Pi,o(z) = 0), we have that degH\(z) is odd. Thus, as had been suggested earlier, H \ is an even-length filter. 4.3.5.4 Analysis Filter Lengths The above results are significant as they show that Ho(z) and H \ ( z ) both have a highly structured form (i.e., their coefficient sequences are each the sum of two highly structured sequences). Examining the expression for deg//rj(z) and d e gH\ (z ) , we can make one further observation regarding the analysis filters. That is, except in the degenerate case in which all of the {^(z)} 2 ^ 1 are identically zero, the analysis filters cannot have the same length. To see why this is so, we proceed as below. Since, by assumption, at least one of the {^(z)} 2 ^ 1 is not identically zero, property 4 of P ( z ) implies that two cases are possible: 1) deg/'o,o(z) > degPo,i(z) and degPi,o(z) > degPi,i(z); or 2) degPo,o(z) < degP0,i(z) and deg Pi fi(z) < degPi;i(z). In the first case, we have from (4.34) and (4.35) that degHo{z) = 1 +2degPo:o(z) and deg//i(z) = 1 + 2 d e g P i t o ( z ) . From property 5 of P(z), however, we know that degPo,o(z) degP^oCz). Therefore, degHo(z) 7^ deg/Y](z). In the second case, we have from (4.34) and (4.35) that deg//0(z) = 1 + 2degPrjj(z) and deg//i(z) = 1 + 2degPi!i(z). From property 5 of P(z), however, we know that degPo,i(z) ^ deg Pi ^ (z). Therefore, deg#o(z) 7^ degH\ (z). By combining the results for the above two cases, we have that deg//o(z) ^ degH\ (z), except in the degenerate case in which all of the {^(z)} 2 ^ 1 are identically zero. Consequently, the analysis filters cannot have the same lengths, except in this degenerate case. 4.3.5.5 Incompleteness of Parameterization In order to belong to the ELASF family, a transform must have analysis filters of the form described above. Obviously, the close relationship between pairs of samples in the analysis filter impulse responses is quite constraining. For this reason, the ELASF family cannot be a complete parameterization of all PR linear-phase FIR ID two-channel filter banks with even-length filters. Furthermore, an even more basic reason exists for this lack of completeness. As noted above, the analysis filters cannot be of equal length, except in the degenerate case in which all of the {^(z)}2^"1 3 1 0 identically zero. 4.3.5.6 Analysis Filter Gains Consider now the DC and Nyquist gains of the analysis filters, Ho and H \ . From property 2 of P ( z ) , we know that Po,i (z) and Pi,o(z) both have antisymmetric coefficient sequences, and consequently, Po,i(* 2)| z=±i=0 and Pi ,o(z 2 ) | z = ±i=0. (4.36) From property 6 of P(z), we have that Po,o(l) — 1 and Pi,i(l) = 1, and consequently, fl)fl(z2)|«=±i = l and P M ( z 2 ) | z = ± i = l . (4.37) Using (4.36) and (4.37), we can deduce from (4.19) that flb(l) = l , flb(-l)=0, # i ( - l ) = - 2 , and #i(l) = 0. Thus, for any transform in the ELASF family, the associated lowpass analysis filter must have DC and Nyquist gains of 1 and 0, respectively, while the associated highpass analysis filter must have Nyquist and DC gains of 2 and 0, respectively. This result is of practical interest, since it is often desirable for a reversible ITI wavelet transform to have a corresponding lowpass analysis filter with a DC gain of 1 (which typically results in a transform with good dynamic range properties). Chapter 4. Nonexpansive Reversible I T I Wavelet Transforms 79 4.3.5.7 Swapping Analysis and Synthesis Filters Consider a filter bank of the form in Figure 4.6 that is constrained to be of the ELASF type as denned by (4.2). The linear version of this filter bank has the canonical form shown in Figure 4.9(a) with the corresponding inverse shown in Figure 4.9(b). Let us denote the analysis filter transfer functions as HQ{Z) and H\ (z) and the corresponding synthesis filter transfer functions as FQ(Z) and Fi(z). Suppose now that we swap the analysis and synthesis filters, allowing the filters to be renormalized in the process. We will show below that this renormalization can always be performed in such a way that the resulting filter bank also belongs to the ELASF family. To date, this fact seems to have been overlooked (e.g., as in [32]). From the diagram, we can see that the analysis polyphase matrix, E ( z ) , is given by 1 0 1 M(z) 0 1 (4.38) Suppose that we now construct a new filter bank with the analysis filters H'Q(z) and H[(z) where H'0(z) = af>zFo(z) and H[ (z) = otizFi (z). In other words, the new analysis filters are chosen to be renormalized versions of the synthesis filters from the original filter bank. Further assume that we continue to employ the same polyphase representation for the new filter bank. Let us denote the new analysis polyphase matrix as E ' ( z ) . From the definition of the polyphase representation, we can show " (E-l(z))TJ. (4.39) Substituting (4.38) in (4.39), we obtain E'(Z): E'(Z) = [°S ° (nU[ ao 0 0 ai 1 0 ~M+l(z) 1 -hk(z) ]) (4.40) l o - r 1 oi .0 IJ u • Suppose now that we choose oto = \ and ai = - 2 . In this case, we can rewrite (4.40) as follows: E " ( z ) =[L°J (n*=> 1 0 -Mz) i 1 0 -Mk+i{z) i • 1 0 1 -A2k{z) 0 1 ) n M J Lo il J = [h](nfc,' [ l o -A2k+l(z) 1 1 0 n o - • [o-L 2 0 1 Lo l 4A 2 t + i(z) 1 i iM(z) 0 1 1 -A2*(z)"h 0 1 \ ) ) 1 0 4Ai(z) 1 Thus, the new analysis polyphase matrix, E ' ( z ) , (corresponding to the swapped filter bank) has the same general form as the original one, E ( z ) . In other words, the new filter bank also has a symmetry-preserving reversible ITI implementation (which belongs to the ELASF family). The above result is practically useful, since, in some cases, the "transposed" filter bank (i.e., the one with the analysis and synthesis filters swapped) may also be effective for coding purposes. 4.3.6 Relationship Between Symmetric Extension and Per-Displace-Step Extension Consider again a system of the form shown in Figure 4.6. In Section 4.3.2, we carefully examined the symmetry prop-erties of the signals {"tH}2^'' {v*M}it=o2' yo[n\> a n d y\["]• In the OLASF case, all of these signals are symmetric, while in the ELASF case, all of these signals are symmetric/antisymmetric, except uo[n], u\[n], and vo[/i]. Since these signals are symmetric/antisymmetric, we can equivalently define symmetric extension in terms of PDS extension (in-troduced in Section 4.2.3) in which case the input to each lifting step filter is extended by symmetric extension (with the appropriate choice of symmetry type and centers). Chapter 4. Nonexpansive Reversible I T I Wavelet Transforms 80 Suppose that we have a filter bank of the form shown in Figure 4.6 that is constrained to be of the OLASF type as defined by (4.1) with < 2 for k = 0,1,... ,2X — 1. In this case, we assert that symmetric extension (with K = 0) is equivalent to constant PDS extension (where the lifting step filter input is extended to the left by repeating its leftmost sample and to the right by repeating its rightmost sample). In the case of both symmetric extension and constant PDS extension, the signals {w* [n] 1 and yo [n] are completely characterized by their samples at indices n = 0,1,..., L^f^J and the signals {^[n]} 2^ 2 a n c * yi [n] are completely characterized by their samples at indices n = 0,1,..., L^T^J- Furthermore, both extension methods yield the same uo[n] and vo[n] for n over their characteristic sample indices. Consider the lifting steps involving the filters {A*} for even k. When filtering with the filter A*, w*[n] only ever requires right extension by one sample (if at all) in order to obtain the value for «i[|_^ 5^ -J]- Suppose first that A' is even. In the case of symmetric extension, the symmetry center of uk\n] at is an odd multiple of \ , so the sample obtained by extension is equal to K^^y 2 - ] . This, however, is the same result obtained by constant PDS extension. Finally, if A7 is odd, one less sample needs to be computed for v*+i [n] than Uk[n], and need not be extended at all. Thus, symmetric extension is equivalent to constant PDS extension for the lifting steps involving the filters {A*} for even k. Consider the lifting steps involving the filters {A*} for odd k. When filtering with the filter A*, v [^«] always requires left extension by one sample in order to obtain the value for v>[— 1]. In the case of symmetric extension, since a symmetry center of v^ fn] is — ^, the sample obtained by extension is equal to v^ fO]. Clearly, this is the same result obtained by constant PDS extension. If N is odd, v*[n] must also be right extended by one sample in order to obtain the value for v^^yi] (since one fewer sample is associated with v*[n] than u^+x [n]). In the case of symmetric extension, the symmetry center is an odd multiple of | , so the sample obtained by extension is equal to v*[^y^]. Again, this is the same result that is obtained from constant PDS extension. Thus, symmetric extension is equivalent to constant PDS extension for the lifting steps involving the filters {A*} for odd k . Combining the above results for both sets of lifting filters, we see that constant PDS extension is equivalent to symmetric extension for the specific case considered. Since constant PDS extension is typically easier to implement than symmetric extension, this equivalence is potentially quite useful. For example, both of the filter banks defined in the JPEG-2000 Part-1 standard (i.e., ISO/IEC 15444-1:2001 [72]) are of the form assumed above. Therefore, one can exploit the equivalence between symmetric extension and constant PDS extension in order to simplify JPEG-2000 codec implementations. For example, this equivalence has been employed by the JasPer software [ 14,76] since at least version 0.044. 4.4 Design of Low-Complexity Symmetry-Preserving Transforms In the preceding sections, we studied two families of symmetry-preserving transforms. These families are of great practical value since their constituent transforms can be used in conjunction with symmetric extension in order to han-dle arbitrary-length signals in a nonexpansive manner. Another important consideration, however, is the computational complexity of transforms. For this reason, we are also interested in low-complexity transforms in these families. In what follows, we employ a simple exhaustive search technique to find effective low-complexity transforms from the two families mentioned above. The transforms obtained are then employed in an image coding system to demonstrate their effectiveness. 4.4.1 Transforms To begin, we recall the form of the OLASF and ELASF families of symmetry-preserving transforms introduced earlier. These families are associated with a UMD filter bank having the general structure shown in Figure 4.6. In the OLASF and ELASF cases, the various transform parameters are chosen as specified by (4.1) and (4.2), respectively. If we neglect the effects of the rounding operations { Q k } , we obtain the linear system shown in Figure 4.9. As a matter of notation, we denote the lowpass and highpass analysis filter transfer functions as HQ(Z) and #t(z), respectively. The number of ladder steps is denoted as N . The kth ladder step filter has transfer function At(z) and length L*. Chapter 4. Nonexpansive Reversible I T I Wavelet Transforms 81 4.4.2 Design Method To date, many criteria have been suggested for UMD filter bank design. For image coding applications, however, the following criteria have proven to be particularly useful: coding gain, analysis filter frequency selectivity, the number of vanishing moments of the analyzing and synthesizing wavelet functions, and the smoothness of the synthesizing scaling and wavelet functions. Moreover, experimental results suggest that UMD filter banks which are effective for image coding generally satisfy the following conditions: \H0(eJ°)\^0 and \H0{e*)\=0, |tf,(<>)|^0 and \Hl(eJ°)\ = 0, (4.41a) fl, > 2 or r\>2, (4.41b) S H o < 0.25 and S H { < 0.25 (4.41c) G 6 > 9 (4.41d) where f3n/8 (|//o(^)l-|//o(^)|)-dco + /5JI/8 SH0= r (\Ho(ej0i)\-\Ho(eJO)\)2d(0+ f \H0{e^)\2d<», (4.42a) SHl= P/*me*a)\2d(0+ [* {\Hx{en\-\Hx{e*)\)2du>, (4.42b) J0 J5n/& G& is the coding gain of the sixfold iterated UMD filter bank (i.e., the coding gain associated with a 6-level wavelet decomposition), and fj and r\ are the number of vanishing moments of the analyzing and synthesizing wavelet func-tions, respectively. The above conditions lead us to propose a simple design algorithm based on an exhaustive search as described below. Suppose that we choose a small finite set of admissible values for the coefficients of the filters {A^}. Given the type of ladder network and its associated parameters, namely the number of ladder steps N , and the filter lengths we can perform an exhaustive search of the solution space, seeking UMD filter banks which satisfy the conditions given by (4.41). Any UMD filter bank satisfying these conditions can be taken as an output of the design process. Clearly, the above design technique has one severe limitation. The size of the solution space must be kept small enough that an exhaustive search is feasible. Hence, we can only use such an approach for small design problems. For our purposes here, however, this limitation is not a serious one. Since we are interested in low-complexity transforms, we wish to keep the number of ladder step filters and their lengths to a minimum. Provided that the number of admissible filter coefficient values is also kept small enough, the resulting design problem is computationally tractable. 4.4.3 Design Examples Our design method was applied to several filter bank configurations. The filter coefficient values were selected to be either powers of two or dyadic rational numbers. Numerous transforms were obtained. Several of the more promising new solutions are listed in Table 4.1. Of these, the 5/11-A and 13/7-A transforms belong to the OLASF family, while the 6/14-A transform belongs to the ELASF family. Note that the 5/11-A and 6/14-A transforms have strictly power-of-two filter coefficients. A number of important parameters for the various transforms are listed in Table 4.2. As can be seen from the table, these transforms have both high coding gain and good frequency selectivity. In addition to the new transforms obtained, our design method also found many of the more performant transforms already described in the literature, including the 5/3, 5/11-C, 9/7-M, 13/7-T, 2/6, and 2/10 transforms discussed in Chapter 5 as well as the 6/14-CRF and 13/7-CRF transforms of [27]. 4.4.4 Coding Results To demonstrate the effectiveness of the new transforms, they were employed in the JPEG-2000 image codec (or, more precisely, version 0.0 of the JPEG-2000 verification model software) [1.12]. Two grayscale images from the JPEG-2000 Chapter 4. Nonexpansive Reversible I T I Wavelet Transforms 82 Table 4.1: Transforms f Mz) = j ( - z - l ) 1(1+^) 5/11-A < { A2{z) = ^ ( z 2 - z - l + z _ l ) ( Mz) = -1 6/14-A Mz) = ^(-z+16+1) { A 2 ( z ) = j2(z2-z + z~l - z ' 2 ) 13/7-A < f Mz) = I A i ( z ) = ± ( S z 2 - \ 9 z x - \ 9 + 3 z ~ l ) ± ( - z + 5 + 5 z - l - z - 2 ) Table 4.2: Transform parameters Transform fi Tl G 6 SH0 SHI 5/11-A 2 2 9.603 0.165 0.131 6/14-A 3 1 9.713 0.075 0.072 13/7-A 2 2 9.729 0.025 0.034 test set [68] (e.g., gold and finger) were used as input data. The lossy and lossless results for the various transforms are given in Tables 4.3 and 4.4, respectively. For comparison purposes, we also include the results obtained with the well-known 5/11-C and 2/10 transforms. Clearly, the new transforms perform quite well for both lossy and lossless compression. In the case of lossy compression, the subjective image quality obtained with the new transforms is also good, as demonstrated by the example depicted in Figure 4.16. In Chapter 5, we also present a detailed performance evaluation of several transforms for image coding purposes, including the 5/11-A transform. As the results of this later evaluation demonstrate, the 5/11-A is very effective for both lossy and lossless coding. 4.5 Summary In this chapter, we studied nonexpansive reversible ITI wavelet transforms. We commenced by introducing various strategies for handling finite-extent signals. Then, we explained how each technique can be used to form nonexpansive transforms. The advantages and disadvantages of the various schemes were discussed. Two families of symmetry-preserving reversible ITI wavelet transforms were studied (i.e., the OLASF and ELASF families). We showed that the transforms from both families are compatible with symmetric extension (as described in Section 4.3.2), and can be used to handle signals of arbitrary length in a nonexpansive manner. The characteristics of the two transform families and their constituent transforms were then studied. For the more constrained of the two families (i.e., the ELASF family), we characterized the transforms belonging to this family. That is, we showed that: 1) such transforms are associated with analysis filters having transfer functions of a highly structured form; 2) the DC and Nyquist gains of the analysis filters are fixed, independent of the choice of lifting step filters; and 3) if a particular filter bank is associated with a transform in the ELASF family, then so too is its "transposed" version. By better understanding the characteristics of this family of transforms, one can hope to better utilize this family in signal coding applications. During the course of our work, we derived some new reversible structures that are useful for the construction of symmetry-preserving transforms. For OLASF systems with length-2 lifting filters, we showed that symmetric extension is equivalent to constant PDS extension. This fact can be exploited in order to reduce the complexity of JPEG-2000 Part-1 codec implementations. A simple exhaustive search technique was explored as a means to design low-complexity symmetry-preserving reversible ITI wavelet transforms for image coding. Several good transforms were obtained. Two of these trans-forms have strictly power-of-two filter coefficients which is attractive for shift-and-add implementations. As lossy and lossless compression results demonstrate, these transforms are all particularly effective, often outperforming the Chapter 4. Nonexpansive Reversible I T I Wavelet Transforms 83 Table 4.3: Lossy compression results for the (a) finger and (b) gold images (a) Bit Rate PSNR (dB) (bpp) 5/11-A 6/14-A 13/7-A 5/11-C 2/10 0.0625 19.93 20.44 20.31 19.90 20.36 0.125 21.52 21.93 21.90 21.55 21.88 0.250 24.15 24.41 24.64 24.27 24.37 0.500 27.66 27.56 28.27 27.91 27.57 1.000 31.51 31.08 32.06 31.77 31.13 2.000 37.52 36.92 37.91 37.77 36.99 (b) Bit Rate PSNR (dB) (bpp) 5/11-A 6/14-A 13/7-A 5/11-C 2/10 0.0625 27.21 27.36 27.33 27.18 27.27 0.125 28.99 29.13 29.12 28.92 29.00 0.250 31.21 31.25 31.26 31.13 31.13 0.500 33.85 33.82 33.85 33.75 33.67 1.000 37.29 37.14 37.16 37.15 36.95 2.000 42.34 42.15 41.85 42.10 42.04 Table 4.4: Lossless compression results Image Bit Rate (bpp) 5/11-A 6/14-A 13/7-A 5/11-C 2/10 finger gold 5.380 4.467 5.457 4.488 5.305 4.483 5.323 4.472 5.415 4.498 Figure 4.16: Part of the gold image, (a) Original. Lossy reconstruction at a bit rate of 0.125 bpp using the (b) 5/11-A, (c) 6/14-A, (d) 13/7-A, (e) 5/11-C, and (f) 2/10 transforms. Chapter 4. Nonexpansive Reversible I T I Wavelet Transforms 85 well-known 5/11-C and 2/10 transforms. The world is a comedy to those who think and a tragedy to those who feel. —Horace Walpole (1717-1797) Chapter 5 86 Revers ib l e I T I Wave le t T r an s f o rms fo r Image C o d i n g If it weren't for the sadness it wouldn't be so bad ... If it were just the rage at my mother's unchanged personality and my father's unchanged silence, then it would be manageable. But my mother per-sists in making small, desperate, pathetic gestures of love; my father persists in attempting an understanding silence. It tears me apart. It makes me want to kill somebody or destroy something while simultaneously committing suicide. The sadness is unbearable, watching my parents try to contain themselves, try not to do anything wrong or alienate me again. Yet their attempts are so obvious, clumsy, revealing all the more starkly what they are trying to hide. Their efforts fall short but are efforts nonetheless, and it makes me feel immensely sad, a sorrow that burns and washes in the biggest fires and oceans. I want to protect them, throw my arms around them and demonstrate total, self-sacrificing love. And I know that I can't because it isn't there. —Evelyn Lau, Runaway: Diary of a Street Kid, 1989 Overview In the context of image coding, a number of reversible ITI wavelet transforms are compared on the basis of their lossy compression performance, lossless compression performance, and computational complexity. Of the transforms considered, several were found to perform particularly well, with the best choice f o r a given application depending on the relative importance of the preceding criteria. Reversible I T I versions of numerous transforms are also compared to their conventional (i.e., non-reversible real-to-real) counterparts for lossy compression. At low bit rates, reversible I T I and conventional versions of transforms were found to often yield results of comparable quality. Factors affecting the compression performance of reversible I T I wavelet transforms are also presented, supported by both experimental data and theoretical arguments. 5.1 Introduction The particular application of reversible ITI wavelet transforms considered in this thesis is image coding. In recent years, reversible ITI wavelet transforms have become a popular tool for image coding applications. Transforms of this type are extremely useful for image compression systems requiring efficient handling of lossless coding, minimal memory usage, or low computational complexity. Furthermore, such transforms are particularly attractive for sup-porting functionalities such as progressive lossy-to-lossless recovery of images, lossy compression with the lossless reproduction of a region of interest, and strictly lossy compression with minimal memory usage. Due to their utility, reversible ITI wavelet transforms have been adopted for use in the JPEG-2000 standard [72,73,19]. In this chapter, several reversible ITI wavelet transforms are compared on the basis of their lossy compression performance, lossless compression performance, and computational complexity. By using the results of such an eval-uation, not only are we able to build systems with improved compression efficiency, but we can also better understand the tradeoffs between compression efficiency and computational complexity. Reversible ITI versions of transforms Chapter 5. Reversible TTI Wavelet Transforms for Image Coding 87 Table 5.1: Transforms Name Description / References 5/3 5/3, (2,2), [55], equation (4.1) in [40] 2/6 2/6,(3, 1), [142], [148] SPB S+P Transform with Predictor B, IIR synthesis filters, [114] 9/7-M 9/7, (4,2), [125], equation (4.2) in [40] 2/10 2/10, (5, 1), [59] 5/11-C 5/11, (4, 2), equation (4.6) in [40] 5/11-A 5/11, (2, 2), [12] 6/14 6/14, (3, 3), [6] SPC S+P Transform with Predictor C, IIR synthesis filters, [ 11.4] 13/7-T 13/7, (4,4), [125], equation (4.5) in [40] 13/7-C 13/7, (4, 2), [6] 9/7-F 9/7, (4,4), [24], section 4.4 in [40] are also compared to their conventional (i.e., non-reversible real-to-real) counterparts for lossy compression. The ob-jective, in this case, is to quantify any performance degradation associated with the introduction of the reversible and ITI properties (especially the latter). If these properties do not adversely affect image quality, this would provide a compelling argument for the use of reversible ITI transforms in strictly lossy compression systems in order to reduce memory and computational requirements. If, however, compression performance is negatively impacted, it would be useful to have some quantitative measure of this degradation. Finally, factors affecting the compression performance of reversible ITI wavelet transforms are discussed, supported by both experimental data and theoretical arguments. Through the insight such information provides, one can hope to design new and more effective transforms. The remainder of this chapter is structured as follows. Section 5.2 begins by introducing the reversible ITI wavelet transforms considered in this study. This is followed, in Section 5.3, by a brief comparison of the various transforms in terms of their computational complexity and memory requirements. Then, in Section 5.4, the transforms are employed for both lossy and lossless coding in a state-of-the-art image compression system and the results compared. Of the transforms considered, several were found to perform particularly well, with the best choice for a given application depending on the relative importance of lossy compression performance, lossless compression performance, and com-putational complexity. At low bit rates, reversible ITI and conventional versions of transforms were found to often yield results of comparable quality. In Section 5.5, the lossy and lossless coding results are analyzed in more detail. Finally, we summarize our results in Section 5.6. Some of the work presented in this chapter has also been published in [16,19,11,12,13,6,9,7,10,8]. 5.2 Transforms The reversible ITI wavelet transforms considered in this study are based on the lifting framework [129], or a general-ization thereof [1,9], and were constructed using the technique described in [40]. In all, twelve transforms known to be effective for image coding were evaluated as listed in Table 5.1. All of these transforms are strictly ID in nature and based on two-channel UMD filter banks. The SPB and SPC transforms are based on filter banks with FIR analysis filters and IIR synthesis filters, not all of which have linear phase. All of the other transforms, however, are based on filter banks with linear-phase FTR filters. In the case of transforms based on FIR filter banks, we use the notation x/y to indicate that the underlying filter bank has lowpass and highpass analysis filters of lengths x and y, respectively. The notation (N,N) is also employed to indicate that the analyzing and synthesizing wavelet functions have N and N vanishing moments, respectively. Since the transforms under consideration are ID in nature, images are handled by transforming the rows and columns in succession. Unlike in the case of conventional (linear) versions of transforms, however, the order in which rows and columns are transformed is important. That is, the inverse transform must operate on rows and columns in the reverse order from that used in the forward transform; otherwise, invertibility cannot be guaranteed. The forward transformation equations for each of the transforms are given in Table 5.2. The input signal, lowpass subband signal, and highpass subband signal are denoted as x[n], s[n], and d[n], respectively. For convenience, we also Chapter 5. Reversible I T I Wavelet Transforms for Image Coding 88 define the quantities so[n] = x[2n] and do[n] = x[2n +1]. The inverse transformation equations can be trivially deduced from the forward transformation equations, and thus are not given here. When transforming finite-length signals, it is necessary to adopt some strategy for handling filtering at the signal boundaries. In image coding applications, symmetric extension [33] is often employed for this purpose. To use symmetric extension, however, the underlying filter bank must preserve signal symmetry. Although most of the transforms considered here are derived from linear filter banks with linear-phase filters, this is not sufficient to ensure compatibility with symmetric extension. Due to the rounding operations introduced in order to force a filter bank to map integers to integers, the symmetry-preserving property is often lost. In our study, one of two techniques for handling finite-length signals is employed, depending on the transform at hand. In both instances, the method employed results in a nonexpansive transform for signals of arbitrary length. In cases where the underlying reversible ITI filter bank preserves signal symmetry, symmetric extension (as described in Section 4.3.2) is used. More specifically, this strategy is employed in the case of the 5/3, 9/7-M, 5/11-C, 5/11-A, 13/7-T, 13/7-C, and 9/7-F transforms. In the case of the other transforms, constant PDS extension (described in Section 4.2.3) is employed. Periodic extension is avoided due to its poor performance in image coding applications. Reversible ITI wavelet transforms approximate their parent linear transforms. For this reason, the characteristics of the parent transforms are of great interest. In the case of linear wavelet transforms, a number of characteristics have been shown to potentially influence a transform's effectiveness for coding purposes, including the number of vanishing moments of the analyzing and synthesizing wavelet functions, the regularity of the analyzing and synthesizing scaling functions, coding gain, and the frequency selectivity of the analysis filters. These parameters are given for each transform in Table 5.3. Also of great importance is the shape of the synthesizing scaling and wavelet functions. It is the shape of these functions that ultimately determines the signals for which a transform is effective. In addition, the shape of these functions determines the nature of the artifacts obtained in lossy signal reconstructions. The synthesizing scaling and wavelet functions for each transform are shown in Figures 5.1-5.12. 5.3 Computational Complexity and Memory Requirements All of the reversible ITI wavelet transforms considered in this study are calculated using only fixed-point arithmetic. In particular, only integer addition/subtraction, multiplication, and division operations are required. Since all filter coefficients are approximated by dyadic rational numbers, all division operations can be implemented as bit shifts. For the purposes of computational complexity analysis, we assume that the two's complement representation for integers is used. By assuming a two's complement representation, we can take advantage of the fact that [JC/2wJ is equivalent to the arithmetic right shift of x by A' bits (as shown much earlier in Proposition 4). Assuming a one-level wavelet decomposition (in one dimension), the number of addition, multiplication, and shift operations required per two input samples for each transform is as given in Table 5.4. In each case, two sets of numbers are provided. The first set shows the number of operations in the straightforward implementation of the transform. The second set, shown in parentheses, indicates the number of operations required if all multiplications are converted to shift and add operations (via Booth's algorithm [30]). This complexity metric is of particular interest for hardware implementations or software implementations on architectures where integer multiplications are more costly than addition and shift operations. In the case where both sets of numbers are the same, only one is given. Symmetry in filter coefficients was considered in the calculation of these values. Note that although a number of the transforms have equal operation counts, some require more multiplications than others. In particular, note that the 5/3, 2/6,5/11-C and 5/11-A transforms are truly multiplierless (i.e., their underlying lifting filters all have coefficients that are strictly powers of two). Evidently, the 5/3 and 2/6 transforms require the least computation, followed by the SPB, 9/7-M, 2/10, 5/11-C, and 5/11-A transforms as a group, and then the 6/14, SPC, 13/7-T, and 13/7-C transforms as a group, and lastly the 9/7-F transform which requires the most computation. As mentioned previously, reversible ITI versions of transforms have the potential for simpler implementations than their conventional counterparts. In the conventional case, all operands are real-valued. In the reversible ITI case, however, many operands are integers. Since, in both cases, approximately the same number of bits are required for the integer part of numbers, it follows that more bits require processing in the conventional case. Thus, reversible ITI transforms can offer reduced computational complexity through a smaller wordsize for arithmetic operations. The amount of memory needed by a transform-based coder can be strongly influenced by the amount of memory required to store transform coefficients. In many image coding systems, transform coefficients are represented as real numbers. Typically, in such systems, 32 bits of memory are needed to store each coefficient. Transforms which Chapter 5. Reversible I T I Wavelet Transforms for Image Coding 89 Table 5.2: Forward transforms 5/3 < ' rf[n]=db[n]-U(Jo[«+l]+*oH)J k s [ n ] = s 0 [ n ] + [ i ( d [ n ] + d [ n - l ] ) + {\ 2/6 < di[/i]=d0[n]-'so[n] , [ n ] = 5 0 [ n ] - r [ U M j w d[n] = + LJ M « + 1] +*[/»- 1]) + £J SPB < =4o[«]-Jo[«] </[«] = di[n] + [ } ( - 3 s [ n + 1] +*[«] +2s[n - 1] +2rfi[fi+ 1]) + 9/7-M < ' d[n] = do[n] + [ X ( ( s 0 [ n + 2]+so[n - 1]) -9(j0[n+ 1] +*o[n])) + 2 J k s [ n ] = s 0 [ n ] + [ i ( d [ n ] + d [ n - l ] ) + \\ 2/10 < d i [ n ] = d o [ n ] - s 0 [ n ] s [ n ] = s 0 [ n ] + [ { d l [ n ] \ k d[n\ = d x [n] + (22(5[n - 1] - s[n + 1]) + 3(s[n + 2]-s[n-2])) + \\ 5/11-C < ' d i [ n } = d o [ n } - [ U s Q [ n + l ) + s 0 [ n } ) \ s [ n ] = s 0 [ n ] + [ i ( d 1 [ n ] + d i [ n - l ] ) + {\ k rf[n] = rfi [n] + Lis ( J l [« + 2] - *i [» + 1] - J i [n] + * i [n - 1]) + ^ J 5/11-A < rfi[n]=db[n]-U(io[H-l] + Jo[n])J 4 n ] = 5 0 [ n ] + L 4 f W W + ^ i [ « - l ] ) + 2 j k d [ n ] = d i [ n ] + l & ( s i [ n + 2 ] - s l [ n + l ] - s l [ n ] + s l [ n - i \ ) + \\ 6/14 < d\[n] = do[n] -so[n] s[n] = s0[n] + [U-dx [n + 1] + [n - 1] + & * i [«]) + k rf[n] = rfj [11] + L i ( * i [« + 2] - [n - 2] + 6 ( - * i [n +1] + sx [n - 1])) + ±J SPC < = d 0 [ n ] - s Q [ n ] s [ n } = s o l n ] + l 2 - d x [ n } \ d [ n } = dx [n] + Li ( -8J [« +1] + 4*[n] + 5s[n - 1] - - 2] + 6di [n + 1]) + ±J 13/7-T < ' d[n) = do[n] + [U{s0[n + 2] + s 0 [ n - 1 ] ) - 9 ( s 0 [ n + 1] + *oM)) + 2 J k s[n] = so[n} + [ l ( - d [ n + l } - d [ n - 2 ] ) + 9(d[n}+d[n - 1])) + ±J 13/7-C < ' d[n] = d0[n] + [Us0[n + 2} + s0[n - 1] - 9 ( s 0 [ n + 1] + s0[n])) + k s[n] = s0[n] + [ l ( 5 ( d [ n ] + d [ n - 1]) - ( d [ n + l]+d[n-2])) + \\ 9/7-F « ' dx[n] = d0[n] + Liig(203(-5o[n+ 1] -so[n])) + £ I sx[n] = s0[n] + | * (217(-rf,[n] - d x \ n - 1])) + ±J d [ n } = d i [ n } + l M m ( s x [ n + l } + s x [ n } ) ) + l 2 - \ { s[n] = 5 , [n] + |.3oW(1817(d, [/»]+*! [n - 1])) + \ \ Chapter 5. Reversible I T I Wavelet Transforms for Image Coding 90 Table 5.3: Transform parameters. The (a) parameter values, and (b) parameter definitions (a) Transform N /V R R Gi G 6 SH0 5/3 2 2 0.000 1.000 6.277 9.579 0.165 0.192 2/6 3 1 0.000 1.000 5.651 9.601 0.134 0.138 SPB 2 1 — — — — 0.134 0.217 9/7-M 4 2 0.142 2.000 6.181 9.566 0.129 0.065 2/10 5 1 0.000 1.804 5.625 9.615 0.134 0.062 5/11-C 4 2 0.000 2.142 6.280 9.550 0.165 0.085 5/11-A 2 2 0.000 1.220 6.300 9.603 0.165 0.131 6/14 3 3 1.000 1.809 5.885 9.750 0.035 0.025 SPC 2 1 — — — — 0.134 0.255 13/7-T 4 4 0.841 2.000 6.241 9.708 0.062 0.065 13/7-C 4 2 0.809 2.000 6.260 9.735 0.029 0.065 9/7-F 4 4 1.034 1.701 5.916 9.883 0.043 0.058 (b) Parameter Definition N number of vanishing moments of analyzing wavelet N number of vanishing moments of synthesizing wavelet R regularity of analyzing scaling function R regularity of synthesizing scaling function Gi coding gain for 1-level decomposition G 6 coding gain for 6-level decomposition frequency selectivity of analysis lowpass filter (as given by (4.42)) frequency selectivity of analysis highpass filter (as given by (4.42)) Table 5.4: Computational complexity Transform Adds Shifts Multiplies Total 5/3 5 2 0 7 2/6 5 2 0 7 SPB 7(8) 4(3) 1(0) 12(11) 9/7-M 8(9) 2(3) KO) 11(12) 2/10 7(10) 2(6) 2(0) 11(16) 5/11-C 10 3 0 13 5/11-A 10 3 0 13 6/14 10(11) 3(5) KO) 14 (16) SPC 8(10) 4(5) 2(0) 14(15) 13/7-T 10(12) 2(4) 2(0) 14 (16) 13/7-C 10(12) 2(4) 2(0) 14 (16) 9/7-F 12 (26) 4(18) 4(0) 20 (44) Chapter 5. Reversible 177 Wavelet Transforms for Image Coding 91 (a) (b) Figure 5.1: Synthesizing scaling and wavelet functions for the 5/3 transform, (a) Scaling function, (b) Wavelet function. (a) (b) (a) (b) Figure 5.2: Synthesizing scaling and wavelet functions for the 2/6 transform, (a) Scaling function, (b) Wavelet function. (a) (b) Figure 5.3: Synthesizing scaling and wavelet func-tions for the 9/7-M transform, (a) Scaling function, (b) Wavelet function. '0 2 4 6 8 10 0 2 4 6 8 (a) (b) Figure 5.5: Synthesizing scaling and wavelet func-tions for the 5/11-C transform, (a) Scaling function, (b) Wavelet function. Figure 5.4: Synthesizing scaling and wavelet functions for the 2/10 transform, (a) Scaling function, (b) Wavelet function. (a) (b) Figure 5.6: Synthesizing scaling and wavelet func-tions for the 5/11-A transform, (a) Scaling function, (b) Wavelet function. Chapter 5. Reversible I T I Wavelet Transforms for Image Coding 92 (a) (b) Figure 5.7: Synthesizing scaling and wavelet functions for the 6/14 transform, (a) Scaling function, (b) Wavelet function. (a) (b) Figure 5.8: Synthesizing scaling and wavelet func-tions for the 13/7-T transform, (a) Scaling function, (b) Wavelet function. (a) (b) (a) (b) Figure 5.9: Synthesizing scaling and wavelet func-tions for the 13/7-C transform, (a) Scaling function, (b) Wavelet function. Figure 5.10: Synthesizing scaling and wavelet functions for the 9/7-F transform, (a) Scaling function, (b) Wavelet function. 0 100 200 300 400 600 0 100 200 300 400 (a) (b) Figure 5.11: Synthesizing scaling and wavelet functions for the SPB transform, (a) Scaling function, (b) Wavelet function. 0 200 400 600 800 0 200 400 600 (a) (b) Figure 5.12: Synthesizing scaling and wavelet functions for the SPC transform, (a) Scaling function, (b) Wavelet function. Chapter 5. Reversible I T I Wavelet Transforms for Image Coding 93 map integers to integers can improve significantly upon this situation. Although some intermediate results are still real-valued, such results are only needed transiently. All stored values, including the final transform coefficients, are integers. In fact, experimental results suggest that for many reversible ITI wavelet transforms, 16-bit integers provide sufficient dynamic range for representing the transform coefficients (at least, for 8-bpp images). For example, in this study, 16-bit integers proved to be sufficient for all of the test images that were 10 bpp or less. By using 16 bits of memory per coefficient instead of 32, memory requirements can be reduced by half. (As an aside, we note that some numerical results showing the dynamic range properties of the 5/3 transform can be found in [19].) One might even be able to use a smaller number of bits to represent transform coefficients in some cases. Using a smaller number of bits, however, does increase the risk of numerical overflow. Fortunately, the invertibility of lifting-based transforms can be made to hold even in the presence of overflow (e.g., by exploiting the properties of modular integer arithmetic as in [44]). Provided that occurrences of overflow are infrequent, lossless compression performance is not seriously affected. In the case of lossy compression, however, one needs to be much more cautious about overflow. Overflow has the potential to cause very disturbing artifacts, especially if the overflow impacts coefficients that affect large areas in the reconstructed image. 5.4 Experimental Results As part of our study, various experimental results were obtained regarding the compression performance of the trans-forms under evaluation. In the sections that follow, we briefly present these results without any analysis. A detailed analysis of the results will be provided in Section 5.5. 5.4.1 Evaluation Methodology For evaluation purposes, the JPEG-2000 verification model (VM) software (versions 0.0 and 2.1) was employed. Although some support for reversible ITI wavelet transforms was provided in this software, much of the functionality required for our analysis was not present. Therefore, in order to facilitate our testing, the original transform-related code was replaced with new and more flexible code. This new transform code is described at some length in [10]. The JPEG-2000 V M software employed in this work handles both lossy and lossless compression using transform-based methods. In the lossy case, a subband transform is applied to the original image, the transform coefficients are quantized using trellis-coded quantization (TCQ) [84], and the resulting TCQ indices are entropy coded to yield the compressed bit stream. In the lossless case, a reversible ITI subband transform is applied to the original image, and the transform coefficients are then bitplane and entropy coded. The various transforms under consideration were employed for decorrelation purposes in the image coder described above, and the results examined. Both lossy and lossless compression were considered in our analysis. Using the various transforms, each of the test images was compressed in a lossy manner at several bit rates (i.e., 0.0625, 0.125, 0.25,0.5,1, and 2 bpp), in addition to being compressed losslessly. In the lossless case, compression performance was measured in terms of the final bit rate. In the lossy case, compression performance was assessed using both the PSNR and subjective image quality metrics. The test data used in our study consisted of the grayscale images listed in Table 5.5, all of which were taken from the JPEG-2000 test set [68]. These images vary considerably in size and cover a wide variety of content. For each set of experimental results, the same number of decomposition levels was employed for all transforms (typically six). Since the goal of the subjective evaluations was only to compare artifacts obtained from the transforms themselves and not tiling artifacts, the use of tiling was avoided whenever possible. More specifically, tiling was used only in the case of the cmpnd2 image due to its large size, where a tile size of 2048 x 2048 was employed. In order to perform the subjective testing, several regions from the various lossy reconstructed images were selected to be scrutinized. Small regions of interest were used (instead of whole images) for two reasons: 1. Many of the test images were too large to be displayed at their original resolution on the computer screen, and displaying the images at reduced resolution hides most of the artifacts introduced by compression. 2. If the whole image is used, there might be too much information for a human observer to analyze, making an accurate and meaningful comparison difficult. Since the human eye often cannot distinguish between a high bit rate lossy reconstruction of an image and the original, the subjective testing was restricted to the lower bit rates under consideration. As most applications typically utilize Chapter 5. Reversible I T I Wavelet Transforms for Image Coding 94 Table 5.5: Test images Image Size Depth Description aerial2 2048 x 2048 8 aerial photograph bike 2048x2560 8 collection of objects cafe 2048x2560 8 outside of cafe cats 3072 x 2048 8 cats chart 1688x2347 8 color chart crapndl 512x768 8 mixed text and graphics cmpnd2 5120x6624 8 mixed text and graphics cr 1744x2048 10 computer radiology ct 512x512 12 computer tomography elev 1201x1201 12 fractal-like pattern finger 512x512 8 fingerprint gold 720x576 8 houses and countryside hotel 720x576 8 hotel mat 1528x1146 8 mountains mri 256x256 11 magnetic resonance s a r i 1024x1024 16 synthetic aperture radar sar2 800x800 12 synthetic aperture radar seismic 512x512 8 seismic data target 512x512 8 patterns and textures tools 1524x1200 8 collection of tools t x t u r l 1024x1024 8 aerial photograph txtur2 1024x1024 8 aerial photograph us 512x448 8 ultrasound water 1465x1999 8 water and land woman 2048x2560 8 lady xray 2048x1680 12 x-ray Chapter 5. Reversible TTI Wavelet Transforms for Image Coding 95 compression ratios from 16:1 to 128:1, the bit rates corresponding to these compression ratios were of most interest. The participants involved in the subjective testing came from diverse backgrounds, some having a strong grounding in image processing and/or image coding, while others were end users with no special technical training. Two types of subjective evaluations were employed in our study as described below. The first type of subjective evaluation served to compare the image quality obtained with different reversible ITI wavelet transforms. For a specific test image and bit rate, a particular region of the lossy reconstructed images obtained with the various transforms was shown under magnification. The original image was also shown for reference purposes. Then, a human observer was asked to assign a simple integer ranking to the different results, with a value of one corresponding to best result and larger values indicating poorer results. In instances where a distinction in quality could not be made between two or more results, these results could be assigned the same rank. This process was repeated for various test images, bit rates, regions of interest within the image, and observers. The rankings were then used to determine the most effective transform from a subjective image quality standpoint. The second type of subjective evaluation was used to compare the quality of the lossy reconstructed images ob-tained with reversible ITI and conventional versions of the same transform. For a specific test image, bit rate, and transform, a particular region of interest in the lossy reconstructed images obtained with the reversible ITI and con-ventional versions of a transform was shown under magnification. Again, the original image was also shown for reference purposes. Then, the observer was asked to rank the two reconstructed images relative to one another. The ranking was done on an integer scale from -2 to 2. The values -2, -1,0, 1,2 correspond to the result obtained with the reversible ITI version of the transform being significantly worse than, slightly worse than, the same as, slightly better than, and significantly better than the result obtained with the conventional version, respectively. This process was repeated for numerous transforms, images, bit rates, and observers. 5.4.2 Lossless Compression Performance Each of the test images in Table 5.5 was compressed in a lossless manner using the various transforms under evaluation, resulting in the bit rates shown in Table 5.6. For each image, the best result has been highlighted. Clearly, no single transform performs best for all of the images. In order to provide additional insight into the results of Table 5.6, relative differences were also computed. For each image, the best transform was used as a reference and the relative differences for the remaining transforms were calculated. Often, the difference between the best and worst transforms is less than 2%. For the images where this difference is more significant (i.e., greater than 2%), the relative performance numbers have been included in Table 5.7. Note that, although the various transforms often yield comparable lossless coding efficiency, their computational complexities can vary significantly, as detailed earlier in Section 5.3. As one might expect, image content is an important factor influencing transform effectiveness. If one considers only natural imagery, some particularly useful observations can be made. In addition to the lossless results for each image, Table 5.6 provides the average bit rate for each transform taken over all of the natural 8-bpp images. For images of this type (e.g., cafe, gold, woman), the 5/11-C, 5/11-A, and 13/7-T transforms generally perform best, followed closely by the 9/7-M and 13/7-C transforms. The 5/3 transform also fares reasonably well considering its very low computational complexity. For images with a relatively greater amount of high-frequency content (e.g., cmpnd2 and target), the 5/3 transform tends to yield the best results, often by a significant margin. In [40] and [39], Calderbank et al. compare the lossless compression performance of several reversible ITI wavelet transforms also considered in our study (i.e., the 2/6, 5/3, SPB, 9/7-M, 5/11-C, and 9/7-F transforms). The image coder employed in their evaluation was based on techniques described in [11.4]. Generally, the results presented by Calderbank and his colleagues coincide with our findings. For example, both studies found that (for lossless compression): 1. The 5/3 transform outperforms the 2/6 transform for the vast majority of images. 2. The 9/7-F transform consistently performs very poorly. 3. The 5/11-C transform fares amongst the best for natural imagery. Interestingly, the 13/7-T transform, which we found to be particularly effective for natural imagery, was described in [40] and [39], but was not included in their performance evaluation. In the context of a reversible embedded image coder based on the EZW coding scheme [119], lossless compression results similar to those above have also been obtained [1,5]. Thus, these results appear to be consistent across most types of lossless coders, especially those employing bitplane coding of the transform coefficients. Chapter 5. Reversible I T I Wavelet Transforms for Image Coding 96 Table 5.6: Lossless compression results Image Bit Rate (bpp) 5/3 2/6 SPB 9/7-M 2/10 5/11-C 5/11-A 6/14 SPC 13/7-T 13/7-C 9/7-F a e r i a l 2 5.243 5.250 5.234 5.221 5.232 5.223 5.226 5.229 5.262 5.217 5.222 5.233 b ike 4.415 4.437 4.417 4.406 4.427 4.407 4.401 4.429 4.454 4.397 4.402 4.456 cafe 5.194 5.215 5.191 5.177 5.203 5.178 5.175 5.217 5.228 5.175 5.186 5.220 cats 2.534 2.528 2.500 2.493 2.499 2.493 2.509 2.493 2.493 2.491 2.494 2.518 chart 3.052 3.095 3.088 3.103 3.115 3.094 3.063 3.176 3.156 3.104 3.117 3.225 cmpndl 2.142 2.143 2.216 2.454 2.368 2.487 2.403 2.578 2.362 2.492 2.523 2.540 cmpnd.2 2.540 2.605 2.655 2.753 2.733 2.770 2.686 2.855 2.790 2.775 2.797 2.906 c r 3.216 3.225 3.230 3.225 3.230 3.225 3.219 3.220 3.262 3.217 3.216 3.221 c t 2.886 2.890 2.739 2.654 2.710 2.663 2.784 2.727 2.578 2.664 2.677 2.778 elev 2.453 2.464 2.408 2.320 2.358 2.317 2.380 2.378 2.350 2.326 2.335 2.530 f lager 5.470 5.550 5.423 5.321 5.415 5.323 5.381 5.380 5312 5.328 5.337 5.372 gold 4.476 4.502 4.497 4.476 4.498 4.471 4.467 4.490 4.539 4.473 4.477 4.518 ho te l 4.458 4.475 4.476 4.458 4.469 4.449 4.440 4.466 4.528 4.448 4.453 4.493 raat 2.994 3.102 3.110 3.036 3.116 2.984 2.983 3.173 3.164 3.049 3.081 3.270 mri 4.022 4.034 3.939 3.926 3.960 3.926 3.970 3.954 3.879 3.925 3.931 3.971 s a r i 12.68 12.69 12.68 12.68 12.69 12.68 12.67 12.69 12.73 12.67 12.68 12.67 sar2 7.610 7.625 7.626 7.615 7.627 7.616 7.610 7.627 7.662 7.612 7.616 7.611 se ismic 2.781 2.807 2.704 2.610 2.684 2.654 2.724 2.653 2.577 2.613 2.615 2.879 target 2.129 2.231 2.234 2.256 2.313 2.247 2.213 2.355 2.242 2.250 2.243 2.546 t oo l s 5.281 5.293 5.272 5.260 5.283 5.263 5.261 5.289 5.308 5.256 5.263 5.286 fcxturl 6.581 6.579 6.590 6.584 6.586 6.586 6.578 6.585 6.644 6.577 6.581 6.578 tx tur2 5.427 5.426 5.430 5.421 5.433 5.425 5.418 5.440 5.486 5.418 5.425 5.435 us 3.031 3.054 3.078 3.169 3.111 3.141 3.104 3.211 3.162 3.167 3.192 3.473 water 3.325 3.334 3.368 3.357 3.354 3.338 3.325 3.333 3.436 3.337 3.331 3.408 woman 4.394 4.407 4.382 4.366 4.387 4.367 4.372 4.385 4.409 4.362 4.368 4.420 x....ray 4.408 4.418 4.408 4.404 4.414 4.407 4.405 4.414 4.435 4.400 4.403 4.410 MeanT 4.599 4.623 4.607 4.583 4.608 4.577 4.580 4.608 4.636 4.579 4.586 4.631 'Mean over natural 8-bpp images (i.e., aerial2, b ike , cafe, ca ts , f i nge r , go ld , h o t e l , mat, t o o l s , tx tu r l , txtuir2, water, woman). Table 5.7: Lossless compression results relative to the best transform for each image Image Increase in Bit Rate (%) 5/3 2/6 SPB 9/7-M 2/10 5/11-C 5/11-A 6/14 SPC 13/7-T 13/7-C 9/7-F chart — 1.4 1.2 1.7 2.1 1.4 0.4 4.1 3.4 1.7 2.1 5.7 cmpndl — 0.0 3.5 14.6 10.6 16.1 12.2 20.4 10.3 16.3 17.8 18.6 cmpnd2 — 2.6 4.5 8.4 7.6 9.1 5.7 12.4 9.8 9.3 10.1 14.4 ct 11.9 12.1 6.2 2.9 5.1 3.3 8.0 5.8 — 3.3 3.8 7.8 elev 5.9 6.3 3.9 0.1 1.8 — 2.7 2.6 1.4 0.4 0.8 9.2 finger 3.0 4.5 2.1 0.2 1.9 0.2 1.3 1.3 — 0.3 0.5 1.1 mat 0.4 4.0 4.3 1.8 4.5 0.0 — 6.4 6.1 2.2 3.3 9.6 mri 3.7 4.0 1.5 1.2 2.1 1.2 2.3 1.9 — 1.2 1.3 2.4 seismic 7.9 8.9 4.9 1.3 4.2 3.0 5.7 2.9 — 1.4 1.5 11.7 target — 4.8 4.9 6.0 8.6 5.5 3.9 10.6 5.3 5.7 5.4 19.6 us — 0.8 1.6 4.6 2.6 3.6 2.4 5.9 4.3 4.5 5.3 14.6 water — 0.3 1.3 1.0 0.9 0.4 — 0.2 3.3 0.4 0.2 2.5 Chapter 5. Reversible I T I Wavelet Transforms for Image Coding 97 5.4.3 PSNR Lossy Compression Performance Each of the test images in Table 5.5 was compressed in a lossy manner at several bit rates using the various transforms under evaluation. The image quality was then measured using the PSNR metric as defined by (2.29). The results for a representative subset of the test images are given in Table 5.8. For each test image and bit rate pair, the best result has been highlighted. For low bit rates (i.e., compression ratios of 16:1 or greater), the 9/7-F, 6/14, 13/7-C, and 13/7-T transforms are consistently the best performers in approximately that order. The SPC and SPB transforms fare the worst. As the bit rate increases, lossy compression performance becomes increasing influenced by the lossless performance characteristics of the transforms. These results will be revisited later in Section 5.5.2 for analysis purposes. In applications where lossy image reconstructions are processed by a computer system (e.g., some scientific ap-plications), the use of a distortion metric based on the mean-squared error (i.e., PSNR) can be quite appropriate. Unfortunately, the PSNR can be a poor measure of image quality for applications in which images are to be viewed by humans. Hence, a subjective evaluation, the results of which are presented in the next section, was necessary. 5.4.4 Subjective Lossy Compression Performance Since the PSNR does not necessarily correlate well with image quality as perceived by the human visual system, we supplemented our PSNR results with a subjective evaluation. This subjective testing was undertaken as specified in Section 5.4.1. In order to make the testing more practical, the evaluation was divided into two stages. The same testing strategy was used in each stage. Only the transforms under evaluation in each case differed. In the first stage, the 5/3,2/6, SPB, 9/7-M, 2/10, SPC, 13/7-T, and 9/7-F transforms were compared. After obtaining the rankings for 14 sets of images using 12 observers, these results were averaged together to yield the overall rankings given in Table 5.9. From these results, we see that the transforms fall into three distinct performance categories. The 9/7-F, 5/3, 13/7-T, and 9/7-M transforms perform best, followed by the 2/6 and 2/10 transforms, followed by the SPB and SPC transforms. These numerical results generally agree with comments made by participants during the subjective tests—namely, that the transforms fall into three distinct performance categories. The second stage of subjective testing included some of the best transforms from the first stage and several new transforms, resulting in the following set of transforms: 5/3, 9/7-M, 5/11-C, 5/11-A, 6/14, 13/7-T, and 13/7-C. After conducting the subjective tests with 14 sets of images and 12 observers, the rankings were averaged together with the results shown in Table 5.10. In this round of tests, the general consensus amongst the test participants is that all of the transforms are very close in terms of subjective quality, with the possible exception of the 6/14 transform which performs slightly worse. In other words, although the average rankings may differ by a large amount in this stage of the testing, it is important to note that even large disparities correspond to very minute differences in subjective image quality. Considering the results from both stages of the subjective testing, we conclude that the 9/7-F, 5/3, 13/7-T, 13/7-C, 5/11-C, 5/11-A, and 9/7-M transforms have the best performance, followed by the 6/14, 2/6, and 2/10 transforms, followed by the SPB and SPC transforms. Examples of the quality obtained with transforms from each performance category are shown in Figure 5.13. We will again revisit these results in Section 5.5.2 where they will be examined in more detail. 5.4.5 Reversible ITI Versus Conventional Transforms for Lossy Compression Using both reversible ITI and conventional versions of the transforms listed in Table 5.1, the test images in Table 5.5 were compressed in a lossy manner at several bit rates. The reconstructed images obtained with the two different versions of each transform were then compared. The difference in the PSNR for a representative subset of the images is given in Table 5.11. A negative value corresponds to poorer performance in the reversible ITI transform case. As is evident from these results, the performance is almost always degraded in terms of PSNR in the reversible ITI transform case, although the degradation is often very small. These numerical results also suggest some other important trends that will be discussed later in Section 5.5.1. Subjective testing was also undertaken to compare the quality of the lossy image reconstructions obtained with reversible ITI and conventional versions of transforms. This testing was performed as described previously in Sec-tion 5.4.1. Due to the time consuming nature of subjective testing, only a subset of the transforms listed in Table 5.1 was used, namely the 5/3, 2/6, SPB, 9/7-M, 2/10, SPC, 13/7-T, and 9/7-F transforms. After averaging the scores Chapter 5. Reversible I T I Wavelet Transforms for Image Coding 98 Table 5.8: Lossy compression results for (a) aerial2, (b) bike, (c) cafe, (d) sar2, and (e) target images (a) Bit Rate PSNR (dB) (bpp) 5/3 2/6 SPB 9/7-M 2/10 5/11-C 5/11-A 6/14 SPC 13/7-T 13/7-C 9/7-F 0.0625 24.40 24.47 24.05 24.46 24.55 24.41 24.45 24.66 23.63 24.60 24.64 24.72 0.1250 26.24 26.31 25.83 26.36 26.42 26.30 26.33 26.53 25.35 26.50 26.54 26.59 0.2500 28.15 28.25 27.49 28.35 28.37 28.26 28.25 28.53 26.91 28.50 28.55 28.60 0.5000 30.36 30.37 29.25 30.49 30.48 30.47 30.45 30.68 28.54 30.62 30.67 30.69 1.0000 33.30 33.30 31.89 33.35 33.31 33.34 33.36 33.54 30.97 33.50 33.53 33.50 2.0000 38.55 38.51 36.74 38.46 38.52 38.45 38.53 38.56 35.72 38.56 38.58 38.13 (b) Bit Rate PSNR (dB) (bpp) 5/3 2/6 SPB 9/7-M 2/10 5/11-C 5/11-A 6/14 SPC 13/7-T 13/7-C 9/7-F 0.0625 22.49 22.63 22.16 22.61 22.70 22.59 22.52 22.88 21.60 22.78 22.85 22.95 0.1250 25.03 25.26 24.63 25.12 25.32 25.03 25.08 25.47 24.02 25.34 25.41 2558 0.2500 28.27 28.34 27.84 28.31 28.37 28.21 28.31 28.56 27.17 28.57 28.62 28.76 0.5000 32.27 32.24 31.59 32.23 32.18 32.13 32.29 32.33 30.77 32.30 32.51 32.57 1.0000 36.97 36.87 35.60 36.85 36.74 36.82 36.95 36.87 34.76 37.04 37.06 36.83 2.0000 42.77 42.57 40.55 42.42 42.43 42.48 42.70 42.35 39.49 42.47 42.42 41.12 (c) Bit Rate PSNR (dB) (bpp) 5/3 2/6 SPB 9/7-M 2/10 5/11-C 5/11-A 6/14 SPC 13/7-T 13/7-C 9/7-F 0.0625 18.77 18.89 18.17 18.78 18.88 18.72 18.77 18.99 17.61 18.92 18.97 19.04 0.1250 20.29 20.47 19.77 20.30 20.45 20.25 20.29 20.59 19.14 20.45 20.50 20.62 0.2500 22.75 22.90 22.10 22.75 22.89 22.72 22.77 23.03 21.39 22.94 22.98 23.12 0.5000 26.33 26.31 25.56 26.36 26.30 26.32 26.39 26.43 24.81 26.55 26.56 26.64 1.0000 31.41 31.30 30.59 31.42 31.26 31.37 31.47 31.38 29.76 31.62 31.62 31.63 2.0000 38.62 38.42 36.98 38.41 38.32 38.39 38.58 38.33 35.98 38.52 38.49 38.06 (d) Bit Rate PSNR (dB) (bpp) 5/3 276 SPB 9/7-M 2710 5/11-C 5/11-A 6/14 SPC 13/7-T 13/7-C 9/7-F 0.0625 22.21 22.33 21.77 22.22 22.34 22.18 22.21 22.41 21.31 22.32 22.34 22.42 0.1250 22.97 23.09 22.28 22.98 23.09 22.94 22.97 23.18 21.76 23.07 23.10 23.18 0.2500 23.99 24.12 23.06 23.95 24.13 23.93 23.97 24.25 22.44 24.09 24.13 24.26 0.5000 25.54 25.63 24.43 25.50 25.60 25.47 25.53 25.80 23.65 25.64 25.70 25.82 1.0000 28.22 28.44 27.08 28.08 28.35 28.12 28.20 28.54 26.03 28.32 28.38 2855 2.0000 33.85 33.86 32.26 33.66 33.74 33.63 33.79 33.98 31.04 33.86 33.90 34.08 (e) Bit Rate PSNR (dB) (bpp) 5/3 276 SPB 9/7-M 2/10 5/11-C 5/11-A 6/14 SPC 13/7-T 13/7-C 9/7-F 0.0625 17.57 17.30 17.04 17.55 17.34 17.46 17.54 18.17 16.41 18.12 18.36 18.95 0.1250 19.89 19.44 18.12 20.20 19.34 19.99 19.98 20.45 17.54 20.73 21.05 20.97 0.2500 23.28 23.01 22.75 23.64 23.18 23.36 23.41 24.76 22.31 24.38 25.48 24.92 0.5000 29.98 29.30 28.81 30.73 29.74 30.64 30.55 31.15 28.49 31.76 3251 31.82 1.0000 41.38 40.03 37.99 41.25 39.79 41.04 40.99 40.68 37.69 41.70 41.94 40.21 2.0000 54.14 54.79 50.79 52.26 52.91 52.07 52.37 52.22 48.42 52.13 51.78 44.60 Table 5.9: Overall subjective rankings from first stage of testing Rank 5/3 2/6 SPB 9/7-M 2/10 SPC 13/7-T 9/7-F 2.51 4.47 6.05 2.97 4.57 6.88 2.88 2.36 Table 5.10: Overall subjective rankings from second stage of testing Rank 5/3 9/7-M 5/11-C 5/11-A 6/14 13/7-T 13/7-C 1.84 2.30 2.34 2.15 2.75 1.77 1.83 Figure 5.13: Part of cafe image, (a) Original. Lossy reconstruction at a bit rate of 0.25 bpp using the (b) 5/3, (c) 2/10, and (d) SPC transforms. Chapter 5. Reversible I T I Wavelet Transforms for Image Coding 100 Table 5.11: Difference in PSNR performance between reversible ITI and conventional versions of transforms for the (a) aerial2, (b) bike, (c) cafe, (d) sar2, and (e) target images (a) . Bit Rate (bpp) PSNR (dB) 5/3 2/6 SPB 9/7-M 2/10 5/11-C 5/11-A 6/14 SPC 13/7-T 13/7-C 9/7-F 0.0625 -0.03 -0.04 -0.54 -0.05 -0.04 -0.03 -0.03 -0.05 -0.69 -0.03 -0.03 -0.03 0.1250 -0.05 -0.06 -0.64 -0.06 -0.07 -0.06 -0.04 -0.06 -0.84 -0.05 -0.02 -0.08 0.2500 -0.14 -0.09 -0.96 -0.10 -0.10 -0.15 -0.07 -0.10 -1.29 -0.09 -0.09 -0.08 0.5000 -0.09 -0.14 -1.31 -0.13 -0.16 -0.13 -0.11 -0.15 -1.63 -0.15 -0.15 -0.16 (b) Bit Rate (bpp) PSNR (dB) 5/3 276 SPB 9/7-M 2/10 5/11-C 5/11-A 6/14 SPC 13/7-T 13/7-C 9/7-F 0.0625 -0.02 -0.02 -0.52 -0.02 -0.03 -0.02 -0.02 -0.02 -0.70 -0.02 -0.02 -0.01 0.1250 -0.04 -0.06 -0.75 -0.05 -0.05 -0.05 -0.05 -0.05 -0.94 -0.04 -0.05 -0.02 0.2500 -0.08 -0.09 -0.71 -0.10 -0.10 -0.10 -0.08 -0.09 -0.87 -0.09 -0.10 -0.07 0.5000 -0.16 -0.21 -0.99 -0.20 -0.23 -0.21 -0.16 -0.22 -1.16 -0.36 -0.22 -0.24 (c) Bit Rate (bpp) PSNR (dB) 5/3 2/6 SPB 9/7-M 2/10 5/11-C 5/11-A 6/14 SPC 13/7-T 13/7-C 9/7-F 0.0625 -0.01 -0.01 -0.62 -0.01 -0.01 -0.01 -0.01 -0.01 -0.84 -0.01 -0.00 -0.01 0.1250 -0.01 -0.02 -0.68 -0.01 -0.02 -0.01 -0.01 -0.01 -0.86 -0.01 -0.01 -0.01 0.2500 -0.02 -0.03 -0.82 -0.04 -0.03 -0.03 -0.03 -0.02 -1.00 -0.03 -0.02 0.02 0.5000 -0.05 -0.06 -0.91 -0.06 -0.07 -0.07 -0.06 -0.07 -0.94 -0.05 -0.07 -0.07 (d) Bit Rate PSNR (dB) (bpp) 5/3 2/6 SPB 9/7-M 2/10 5/11-C 5/11-A 6/14 SPC 13/7-T 13/7-C 9/7-F 0.0625 -0.01 -0.00 -0.54 -0.00 -0.00 0.00 -0.00 0.00 -0.79 -0.00 0.00 -0.00 0.1250 -0.00 -0.00 -0.76 -0.00 -0.00 0.00 -0.00 -0.00 -1.05 -0.00 -0.00 -0.00 0.2500 0.03 -0.00 -0.97 0.00 -0.00 -0.00 0.00 -0.00 -1.24 0.00 -0.00 0.00 0.5000 -0.00 -0.03 -1.10 -0.00 -0.00 -0.00 -0.00 -0.00 -1.37 -0.00 -0.00 0.00 1.0000 -0.00 0.00 -1.15 -0.07 0.00 -0.00 -0.00 -0.00 -1.36 -0.00 0.07 -0.00 2.0000 -0.00 -0.01 -1.27 -0.00 -0.00 -0.00 -0.01 -0.00 -1.39 -0.00 0.00 0.00 (e) Bit Rate PSNR (dB) (bpp) 5/3 276 SPB 9/7-M 2/10 5/11-C 5/11-A 6/14 SPC 13/7-T 13/7-C 9/7-F 0.0625 0.01 -0.05 -0.34 -0.01 -0.03 0.01 0.02 0.08 -0.69 -0.01 -0.01 0.05 0.1250 0.00 0.03 -1.01 -0.00 -0.14 -0.03 -0.02 -0.06 -1.12 -0.00 -0.11 -0.07 0.2500 0.00 -0.35 -0.28 -0.02 -0.12 -0.03 -0.03 -0.01 -0.87 -0.02 -0.03 -0.03 0.5000 -0.05 -0.25 -1.20 -0.11 -0.14 -0.14 -0.13 -0.11 -1.55 -0.20 -0.17 -0.33 Chapter 5. Reversible I T I Wavelet Transforms for Image Coding 101 Table 5.12: Difference in subjective performance between reversible ITI and conventional versions of transforms Rank 5/3 2/6 SPB 9/7-M 2/10 SPC 13/7-T 9/7-F -0.20 -0.41 -0.80 -0.29 -0.36 -0.77 -0.37 -0.24 obtained using 7 sets of images (per transform) and 10 observers, the results in Table 5.12 were obtained. These numerical results indicate the relative performance of the reversible ITI and conventional versions of the transforms. A negative value corresponds to better performance in the conventional transform case. From the experimental results, we can see that in most cases there is only a small bias favoring the conventional version of a transform. Thus, there is a very slight performance penalty associated with using reversible ITI transforms in place of their conventional counterparts. In many practical applications, however, this difference in performance is not likely to be of concern. Again, we will revisit the results for analysis purposes in Section 5.5.1. 5.5 Analysis of Experimental Results Having briefly presented the various experimental results, we now examine these results in more detail in the sections that follow. To begin, we discuss the factors that affect the relative lossy compression performance of reversible ITI and conventional versions of transforms. Then, factors influencing the compression performance of reversible ITI wavelet transforms are presented. 5.5.1 Reversible ITI Versus Conventional Transforms for Lossy Compression In Section 5.4.5, we presented results comparing the lossy compression performance of reversible ITI and conventional versions of transforms. Examining these results more closely, we see that several factors affect the performance of reversible ITI transforms relative to their conventional counterparts: 1) whether the underlying filter bank has IIR filters, 2) the number of lifting steps in the filter bank, 3) the rounding function used in the lifting steps, 4) the depth of the image (in bpp) being compressed, and 5) the bit rate used for compression of the image. Each of these factors is discussed in detail below. 5.5.1.1 Impact of IIR Filters In the calculation of reversible ITI wavelet transforms, results are rounded to integer values in numerous places. If the underlying filter bank has IIR filters, this rounding occurs inside feedback paths. Thus, rounding errors can accumulate indefinitely1. For this reason, there is a much bigger difference between the reversible ITI and conventional versions of a transform in the case where the transform is associated with a filter bank having IIR filters. In the case of the transforms studied, the SPB and SPC transforms are associated with such filter banks. It is evident that these two transforms suffer the greatest degradation in performance as a result of the reversible and ITI properties. The impact of having IIR filters is illustrated by Figures 5.14 and 5.15. Figures 5.14(a) and 5.14(b) show typical results in the case of transforms based on FIR filter banks. In this case, there is little perceptible difference between the results obtained with the reversible ITI and conventional versions of the same transform. Figures 5.15(a) and 5.15(b) demonstrate typical results obtained in the case of transforms based on filter banks with IIR filters. In this case, the reversible ITI version of the transform yields much poorer results due to increased ringing. 5.5.1.2 Number of Lifting Steps Clearly, the number of lifting steps also influences the compression performance. Since each lifting step further causes the approximation error to increase, transforms with fewer lifting steps will tend to perform better, all other things being equal. Consider, for example, the 5/3, 5/11-C, and 5/11-A transforms. All three transforms have the same first two lifting steps. The 5/3 transform, however, has only two lifting steps, while the other two transforms have three. At least in terms of PSNR performance, the 5/3 transform typically has less degradation in performance than the 5/11-C and 5/11-A transforms. 'Moreover, it may even be difficult to show that the resulting nonlinear system is stable. Chapter 5. Reversible ITI Wavelet Transforms for Image Coding 102 Figure 5.15: Part of the cafe image. Lossy reconstructed image at a bit rate of 0.25 bpp using the (a) conventional and (b) reversible ITI versions of the SPC transform. Chapter 5. Reversible I T I Wavelet Transforms for Image Coding 103 5.5.1.3 Rounding Function The function used to round results to integer values also affects the difference in performance between reversible ITI transforms and their conventional counterparts. In this study, we chose to use a biased floor function (i.e., Q(x) = lx ~*~ ?J) f° r t m s P u r P o s e - O n e could conceivably remove the bias of one-half, however. Such a change would reduce computational complexity at the expense of introducing systematic error. The resulting performance degradation at low bit rates is typically small. At higher bit rates, however, the degradation can be more significant, especially in the case of lossless compression. 5.5.1.4 Depth of Image As the dynamic range of the signal grows, the errors introduced by rounding various results to integer values are relatively smaller. For 8-bpp images there is a clear difference in PSNR performance between the reversible ITI and conventional versions of a transform, especially at higher bit rates. In the case of 12-bpp images (e.g., sar2), however, there is little difference in performance provided that the transforms are based on FIR filter banks. 5.5.1.5 Bit Rate As the bit rate is lowered, the difference in compression performance between reversible ITI and conventional versions of a transform decreases. At a sufficiently low bit rate, the quantization of transform coefficients becomes so coarse that any errors introduced by rounding intermediate results to integers are masked by quantization of the coefficients themselves. 5.5.2 Factors Affecting Compression Performance In our study, three factors were found to affect the compression performance of a reversible ITI wavelet transform: 1) the performance characteristics of its parent linear transform, 2) how closely it is able to approximate its parent linear transform, and 3) the dynamic range of the coefficients produced by the transform. The first two factors affect both lossy and lossless compression performance. The last factor is typically only important in the case of lossless compression performance. Each of these factors will be discussed in turn in the remainder of this section. Table 5.14 lists several parameters for each transform relevant to compression performance. Comments about these parameters will also be made at the appropriate point in the paragraphs that follow. 5.5.2.1 Parent Linear Transform The first point above has been covered in depth by many researchers, and techniques for the design of linear wavelet transforms are plentiful in the literature. Some design considerations include: time and frequency resolution, number of vanishing moments, regularity, and coding gain. The effectiveness of a particular transform is also to some extent signal dependent. Section 5.4.3 presented the PSNR lossy compression performance results for the various transforms. The 9/7-F, 6/14, 13/7-C, and 13/7-T transforms were found to be most effective. Looking at parameter in Table 5.3, we can see that these four transforms also have the parent linear transforms with the highest coding gain. Thus, the coding gain of the parent linear transform is of great importance, as one might suspect. In Section 5.4.4, the subjective lossy compression performance results for the various transforms were presented. To a large extent, the subjective performance of these transforms can be explained by examining the synthesizing scaling and wavelet functions of their parent linear transforms as shown in Figures 5.1-5.12. From these plots, we can see that the 9/7-F, 13/7-T, 13/7-C, 5/11-C, 5/11-A, and 9/7-M transforms have very smooth synthesizing scaling and wavelet functions, leading to their favorable performance. In the case of the 5/3 transform, the scaling and wavelet functions have discontinuities in their first derivatives. Since there are no oscillations or jump discontinuities in these functions, however, the 5/3 transform is still able to perform very well. In fact, the 5/3 transform is especially attractive as it often introduces less pronounced ringing artifacts than the other transforms due to its short filters. Although sometimes very good at preserving high frequency textures, the SPB and SPC transforms perform worst, due largely to their tendency to introduce ringing artifacts. This behavior can be attributed to the fact that both transforms are associated with filter banks having IIR synthesis filters. The 6/14 transform often introduces more ringing artifacts than some of the other transforms, leading to its relatively poor performance. This ringing is associated with the greater oscillatory nature of the synthesizing wavelet and scaling functions. In the case of the 2/6 transform, blockiness often results. This behavior can be attributed to the relatively rougher synthesizing scaling and wavelet functions associated with this transform. Lastly, the 2/10 transform often results in more pronounced ringing artifacts, leading to its poorer performance. This behavior can be attributed to the larger support of the synthesizing scaling and wavelet functions. Chapter 5. Reversible TTI Wavelet Transforms for Image Coding 104 Table 5.13: Influence of the number of lifting steps on lossless compression performance 2 lifting steps 4 lifting steps Bit Rate Bit Rate Image M A E t (bpp) M A E t (bpp) gold 0.409 4.473 0.540 4.479 target 0.216 2.250 0.295 2.311 woman 0.415 4.362 0.536 4.373 mean absolute error 5.5.2.2 Approximation Behavior The approximation characteristics of reversible ITI wavelet transforms have not been given much serious attention to date. This is, however, an important factor to consider. Clearly, even if a reversible ITI transform is derived from a linear transform with desirable properties, the extent to which these properties are present in the reversible ITI transform depends entirely on how well it approximates its parent transform. The predominant source of error is the rounding of intermediate results to integers. This error depends on the rounding function employed, the number of lifting steps in which this rounding is performed, and whether recursive filtering structures are used. The number of lifting steps associated with each transform is given in Table 5.14. In the case of lossless com-pression, most of the best transforms tend to have only two or three lifting steps. Al l other things being equal, as the number of lifting steps increases so too does the approximation error. In the lossy case, the approximation error is less critical. Unless the approximation error is very large, it is usually masked by the error introduced by transform coeffi-cient quantization. Notice in particular that the 9/7-F transform has the most lifting steps of the transforms considered, and it also has arguably the poorest lossless compression performance. To further demonstrate the influence of the number of lifting steps on compression performance, we losslessly compressed several images using two different versions of the 13/7-T transform. The first version of the transform is the one given in Table 5.2 which has two lifting steps. The second version has four lifting steps. In each case, we measured the lossless bit rate and average approximation error. As can be seen from the results in Table 5.13, the lossless compression performance is better for the version of the transform with two lifting steps. Note that the approximation error is also lower in this case as well. This observation suggests that, all other things being equal, transforms with two lifting steps (i.e., interpolating transforms) may offer the most promise for lossless compression2. 5.5.2.3 Dynamic Range In the case of lossless compression, all bits of the transform coefficients must be encoded. Thus, as the dynamic range of the transform coefficients increases, so too does the number of bits required to encode these coefficients. For this reason, dynamic range is an important factor affecting lossless compression performance. Typically, the coefficients obtained by the application of a transform have a greater dynamic range than the original samples. This dynamic range growth is, therefore, of concern. The worst-case dynamic range growth of a particular transform is an approximate function of the 1-norm of its analysis filters. The 1-norms of the lowpass and highpass analysis filters determine the worst-case growth in the lowpass and highpass channels, respectively. As the 1-norms increase, the worst-case dynamic range growth also increases. For reference purposes in what follows, the analysis filter 1-norms for each transform are given in Table 5.14. Since, in the case of a wavelet transform, the lowpass subband signal is successively decomposed, the worst-case dynamic range growth associated with the lowpass channel can potentially have the most effect on the dynamic range of the transform coefficients. To date, much attention has been focused on worst-case lowpass-channel dynamic range growth. For example, it is often cited that the 2/6 and 2/10 transforms are particularly well suited to lossless compression as they have no dynamic range growth in the lowpass channel. While the no lowpass-channel growth attribute can help contribute to good lossless compression performance, it is not good to place too much emphasis on this property. The first problem with the above philosophy is that it is based on a worst-case argument. The class of signals that lead to transform coefficients having the maximum possible magnitudes are rarely encountered in practice. The worst case and actual dynamic ranges are thus quite different. Therefore, one must instead consider the typical distribution 2 Incidentally, using similar reasoning, it would seem that reversible ITI M-band wavelet transforms (where M > 2) [1] are not likely to be as effective for lossless coding as two-band transforms. Chapter 5. Reversible U I Wavelet Transforms for Image Coding 105 LL2 HL2 HL\ HH2 HLo LHi HH\ LH0 HH0 Figure 5.16: Subband structure for the 2D wavelet transform formed by two ID transforms (three-level decomposi-tion). of coefficient magnitudes as opposed to the worst-case scenario. There is, however, a more fundamental problem with placing too much emphasis on lowpass-channel dynamic range growth. Namely, reduced growth in the lowpass chan-nel often comes at the expense of increased growth in the highpass channel. For images, however, the ID transform is applied twice at each level of wavelet decomposition: once horizontally and once vertically. This yields a subband structure like that shown in Figure 5.16. This means that one quarter of all transform coefficients (i.e., the HHo band) are the result of two highpass filtering stages. Similarly, three quarters of all coefficients have been obtained by high-pass filtering in at least one direction. Thus, if the lowpass-channel growth is made smaller at the cost of increased highpass-channel growth, this can, at some point, become counterproductive. Although the lower frequency bands may have coefficients with a smaller dynamic range, the number of coefficients in these bands is relatively small, and their influence is less significant. Consider, for a moment, the 5/3 and 2/6 transforms, both being comparable in complexity. Unlike the, 5/3 trans-form, the 2/6 transform has no growth in the lowpass channel. The 5/3 transform, however, has a lower highpass-channel growth. (See Table 5.14.) In Section 5.4.2, the 5/3 transform was shown to outperform the 2/6 transform for lossless compression on the vast majority of the test images. In Table 5.15, for three of the test images, we show the average coefficient magnitude for the HLo, LHo, and HHo bands individually and all bands together, and the lossless bit rate obtained with each transform. For the gold and target images, where the 5/3 transform yields a notably lower bit rate, the HLo, LHo, and HHo bands have coefficients with lower average magnitudes in the 5/3 case. The average coefficient magnitude is also lower for both images with the 5/3 transform. In the case of the cmpndl image, where both transforms yield approximately the same bit rate, the average coefficient magnitudes for the HLo, LHo, and HHo bands are much closer. Since the coefficient magnitudes are closer in this instance, the lower average magnitudes for the lower frequency bands in the 2/6 transform case become a dominating factor influencing performance. From the above argument, we can conclude that the amount of growth in the lowpass channel is not necessarily a good predictor of a transform's effectiveness for lossless compression. In fact, in Section 5.4.2, we saw that the best transforms for lossless compression all have worst-case growth in the lowpass channel. Clearly, one must also consider the amount of growth in the highpass channel. Examining the 13/7-T and 13/7-C transforms, we observe that they have very similar lifting filter coefficients. The 13/7-T transform, however, has analysis filters with smaller 1-norms. In Section 5.4.2, the 13/7-T transform was found to often outperform the 13/7-C transform for lossless compression. Again, this helps to illustrate that, all other things being equal, the transform with analysis filters having smaller 1-norms will typically perform better for lossless compression. In the case of the 9/7-F transform, another factor helps to contribute to its poor lossless compression performance. This transform is the only transform of those considered with a lowpass analysis filter having a DC gain greater than one. (See Table 5.14.) This larger gain contributes to increased growth in the dynamic range of the transform coefficients. Chapter 5. Reversible TTI Wavelet Transforms for Image Coding 106 Table 5.14: Parameters of transforms affecting compression performance #of Lifting \Ho(eJO)\ |ff,(e*)| Transform Steps INh HAilli 5/3 2 1.500 2.000 1.000 2.000 2/6 3 1.000 2.500 1.000 2.000 SPB 3 1.000 2.750 1.000 2.500 9/7-M 2 1.500 2.250 1.000 2.000 2/10 3 1.000 2.781 1.000 2.000 5/11-C 3 1.500 2.219 1.000 2.000 5/11-A 3 1.500 2.109 1.000 2.000 6/14 3 1.250 2.797 1.000 2.000 SPC 3 1.000 3.125 1.000 2.750 13/7-T 2 1.625 2.250 1.000 2.000 13/7-C 2 1.750 2.250 1.000 2.000 9/7-F 4 1.698 2.109 1.230 1.625 Table 5.15: Average coefficient magnitudes and lossless bit rates for the 5/3 and 2/6 transforms in the case of the (a) gold, (b) target, and (c) cmpndl images (a) Avg. Coef. Magnitude Bit Rate Transform LHo HLo HHo All (bpp) 5/3 4.489 4.259 3.785 5.056 4.476 2/6 4.841 4.796 5.488 5.287 4.502 (b) Avg. Coef. Magnitude Bit Rate Transform LHo HLo HHo All (bpp) 5/3 15.951 15.874 12.165 15.030 2.129 2/6 18.073 18.206 19.705 17.312 2.231 (c) Avg. Coef. Magnitude Bit Rate Transform LHo HLo HHo All (bpp) 5/3 8.834 7.695 7.218 8.887 2.142 2/6 8.610 7.788 9.836 8.641 2.143 Chapter 5. Reversible TTI Wavelet Transforms for Image Coding 107 5.6 Summary Several reversible ITI wavelet transforms have been compared on the basis of their lossy compression performance, lossless compression performance, and computational complexity. For lossless compression, image content is an important factor influencing transform effectiveness. For smooth images, the 5/11-C, 5/11-A, and 13/7-T transforms are most effective, while the 5/3 transform performs best for images with a greater amount of high frequency content. In the case of lossy compression, both PSNR and subjective performance measures were employed. In terms of PSNR performance, the 9/7-F, 6/14, 13/7-C, and 13/7-T transforms yield the best results. At the same time, the 9/7-F, 5/3, 9/7-M, 5/11-C, 5/11-A, 13/7-T, and 13/7-C transforms achieve the most favorable subjective ratings. Of the transforms considered, the 5/3 and 2/6 transforms have the lowest computational complexity while the 9/7-F transform has the highest. Unfortunately, the 9/7-F transform, although very good for lossy compression, is significantly higher in complexity than all of the other transforms, and performs poorly for lossless compression. Obviously, no one transform has superior lossy and lossless compression performance for all classes of images in addition to low computational complexity. Therefore, tradeoffs are involved, and the most appropriate choice of transform depends on the needs of the application at hand. For example, in the case of applications requiring low computational complexity, the 5/3 transform is particularly attractive. It performs reasonably well for both lossy and lossless compression, and has the lowest computational complexity of all of the transforms evaluated. Reversible ITI and conventional versions of the same transforms were also compared. Using reversible ITI versions of transforms instead of their conventional counterparts for lossy compression was found to only introduce a slight degradation in subjective image quality in many cases. When transforms associated with FIR filter banks are employed, the performance penalty is almost negligible. At low bit rates, the difference in PSNR performance is also small. By examining the performance characteristics of the various reversible ITI wavelet transforms, we were also able to determine several factors that affect the compression performance of such transforms. By considering these factors in the future, we can hope to design new and more effective transforms. Sit down before fact as a little child, be prepared to give up every preconceived notion, follow humbly wherever and to whatever abysses nature leads, or you shall learn nothing. —Thomas H. Huxley (1825-1895) Chapter 6 Conc lu s i ons a n d F u t u r e Resea r ch 108 Time passes in moments, moments which, rushing past, define the path of a life just as surely as they lead towards its end. How rarely do we stop to examine that path, to see the reasons why all things happen, to consider whether the path we take in life is our own making or simply one into which we drift with eyes closed? But what if we could stop, pause to take stock of each precious moment before it passes? Might we then see the endless forks in the road that have shaped a life? And, seeing those choices, choose another path? —Dana Scully, The X-Files: All Things 6.1 Conclusions In this thesis, we have studied reversible ITI wavelet transforms in the context of image coding. In so doing, we have helped to advance existing knowledge in this area. In our work, we proposed the GRITIT framework, a single unified framework for reversible ITI wavelet/block transforms. This new framework was then used to study several previously developed frameworks and their inter-relationships. In so doing, we were able to gain some interesting new insights into the various frameworks under consideration. For example, the ORT framework was shown to be a special case of the lifting framework with only trivial extensions. Rounding operators were examined in some detail. Several common rounding operators were characterized in terms of their approximation properties and computational complexity, and relationships between these rounding oper-ators were also examined. We showed that, when one considers approximation error and/or computational complexity, the floor, biased floor, and biased ceiling functions are arguably the most practically useful rounding operators of those studied. The GRITIT framework was demonstrated to be useful for the construction of reversible ITI block transforms. The GST, a family of reversible ITI block transforms, was proposed. Then, the GST was studied in detail, leading to a number of interesting insights about transforms in this family. Moreover, the GST framework was used as a means to derive reduced complexity implementations of the S transform and RCT. A new transform belonging to the GST family, called the MRCT, was introduced, and shown to be practically useful for image coding applications. Strategies for handling the transformation of arbitrary-length signals in a nonexpansive manner were considered (e.g., periodic extension, symmetric extension, and per-displace-step extension). Two families of symmetry-preserving transforms (which are compatible with symmetric extension) were introduced and studied. We characterized the transforms belonging to these families. Some new reversible ITI structures that are useful for constructing symmetry-preserving transforms were also proposed. A simple search-based design technique was explored as means for finding effective low-complexity transforms in the above-mentioned families, and shown to produce good results. In the context of image coding, a number of reversible ITI wavelet transforms were compared on the basis of their lossy compression performance, lossless compression performance, and computational complexity. Of the transforms considered, several were found to perform particularly well, with the best choice for a given application depending on the relative importance of the preceding criteria. Reversible ITI versions of numerous transforms are also compared to their conventional (i.e., non-reversible real-to-real) counterparts for lossy compression. At low bit rates, reversible ITI and conventional versions of transforms were found to often yield results of comparable quality. Factors affecting the compression performance of reversible ITI wavelet transforms were also presented, supported by both experimental data and theoretical arguments. Chapter 6. Conclusions and Future Research 109 6.2 Future Research Although this thesis has made numerous contributions concerning reversible ITI wavelet transforms and their applica-tion to image coding, further work in this area would still be beneficial. Some future research directions are discussed briefly below. As described in this thesis, the GRITIT framework can be used to construct (multidimensional) reversible ITI wavelet transforms that are based on nonseparable sampling and filtering operations. This said, however, our work has focused primarily on the separable case. Since there may be potential benefits offered by transforms based on nonseparable operations, a more thorough consideration of the nonseparable case would certainly be a worthwhile endeavor. In this thesis, we have focused exclusively on non-adaptive transforms. Obviously, adaptive transforms have the potential to be much more effective for signals with highly non-stationary statistics. Thus, work in the area of adaptive transforms could have significant benefits. Obviously, much work could be done in devising good adaptation schemes. Also, clever new techniques for handling the synchronization of the encoder and decoder adaptation algorithms would be quite useful. In our work, we have considered only reversible ITI transforms that approximate linear wavelet/block transforms. One might want to further explore reversible ITI transforms that are based on operations other than linear filtering (e.g., median filtering, morphological operators). Such generalizations are easily accommodated by the GRITIT framework, as discussed earlier. The reversible ITI transform frameworks described'in this thesis can be applied to both wavelet and more general subband decompositions. In our work, we have not considered the general subband case explicitly. There may be potential benefits from doing so. Thus, this presents is another possible direction for future research. 6.3 Closing Remarks This thesis has studied reversible ITI wavelet/block transforms and their application to image coding. By exploiting the research results presented herein, we can hope to find more effective reversible ITI wavelet/block transforms, and build image compression systems with improved coding efficiency. Consequently, the research documented in this thesis can benefit the many applications in which images are stored and communicated. In this universe the night was falling; the shadows were lengthening towards an east that would not know another dawn. But elsewhere the stars were still young and the light of morning lingered; and along the path he once had followed, Man would one day go again. —Arthur C. Clarke, The City and the Stars, 1956 110 B i b l i o g r a phy [1] M . D. Adams. Reversible wavelet transforms and their application to embedded image compression. M.A.Sc. thesis, Department of Electrical and Computer Engineering, University of Victoria, Victoria, BC, Canada, Jan-uary 1998. [2] M . D. Adams. The official JasPer home page, http: / /www. ece.ubc.ca/~mdadams/jasper, 2000. [3] M . D. Adams. JasPer software reference manual. ISO/TEC JTC 1/SC 29/WG 1 N 2415, December 2001. [4] M . D. Adams. The JPEG-2000 still image compression standard. ISO/TEC JTC 1/SC 29/WG 1 N 2412, September 2001. Available online from http: / /www.ece.ubc. ca/~mdadams and distributed with the JasPer software. [5] M . D. Adams and A. Antoniou. Reversible EZW-based image compression using best-transform selection and selective partial embedding. IEEE Trans, on Circuits and Systems—//: Analog and Digital Signal Processing, 47(10): 1119-1122, October 2000. [6] M . D. Adams, I. Kharitonenko, and F. Kossentini. Report on core experiment CodEff4: Performance evaluation of several reversible integer-to-integer wavelet transforms in the JPEG-2000 verification model (version 2.1). ISO/TEC JTC 1/SC 29/WG 1 N 1015, October 1998. [7] M . D. Adams and F. Kossentini. Performance evaluation of different decorrelating transforms in the JPEG-2000 baseline system. In Proc. of IEEE Symposium on Advances in Digital Filtering and, Signal Processing, pages 20-24, Victoria, BC, Canada, June 1998. [8] M . D. Adams and F. Kossentini. Performance evaluation of different reversible decorrelating transforms in the JPEG-2000 baseline system. ISO/TEC JTC 1/SC 29/WG 1 N 750, March 1998. [9] M . D. Adams and F. Kossentini. Performance evaluation of several reversible integer-to-integer wavelet trans-forms in the JPEG-2000 baseline system (VM release 0.0). ISO/TEC JTC 1/SC 29/WG 1 N 866, June 1998. [10] M . D. Adams and F. Kossentini. SBTLIB: A flexible computation engine for subband transforms. ISO/TEC JTC 1/SC 29/WG 1 N 867, June 1998. [11] M . D. Adams and F. Kossentini. Evaluation of reversible integer-to-integer wavelet transforms for image com-pression. In Proc. of IEEE International Conference on Image Processing, volume 3, pages 541-545, Kobe, Japan, October 1999. [12] M . D. Adams and F. Kossentini. Low-complexity reversible integer-to-integer wavelet transforms for image coding. In Proc. of IEEE Pacific Rim Conference on Communications, Computers and Signal Processing, pages 177-180, Victoria, BC, Canada, August 1999. [13] M . D. Adams and F. Kossentini. Performance evaluation of reversible integer-to-integer wavelet transforms for image compression. In Proc. of IEEE Data Compression Conference, page 514, Snowbird, UT, USA, March 1999. [14] M . D. Adams and F. Kossentini. JasPer: A software-based JPEG-2000 codec implementation. In Proc. of IEEE International Conference on Image Processing, volume 2, pages 53-56, Vancouver, BC, Canada, October 2000. [15] M . D. Adams and F. Kossentini. On the relationship between the overlapping rounding transform and lifting frameworks for reversible subband transforms. IEEE Trans, on Signal Processing, 48(l):261-266, January 2000. Bibliography 111 [16] M . D. Adams and F. Kossentini. Reversible integer-to-integer wavelet transforms for image compression: Performance evaluation and analysis. IEEE Trans, on Image Processing, 9(6): 1010-1024, June 2000. [17] M . D. Adams and F. Kossentini. Generalized S transform. In Proc. o f IEEE International Conference on Acoustics, Speech, and Signal Processing, volume 3, pages 1749-1752, Salt Lake City, UT, USA, May 2001. [18] M . D. Adams, F. Kossentini, and R. Ward. Generalized S transform. IEEE Trans, on Signal Processing, 2002. To appear. [19] M . D. Adams and R. Ward. Wavelet transforms in the JPEG-2000 standard. In Proc. of IEEE Pacific Rim Conference on Communications, Computers and Signal Processing, volume 1, pages 160-163, Victoria, BC, Canada, 2001. [20] M . D. Adams and R. Ward. Symmetry-preserving reversible integer-to-integer wavelet transforms. In Proc. of IEEE International Conference on Acoustics, Speech, and Signal Processing, volume 3, pages 2509-2512, Orlando, FL, USA, May 2002. [21] M . D. Adams and R. Ward. Symmetry-preserving reversible integer-to-integer wavelet transforms. Submitted to IEEE Trans, on Signal Processing, January 2002. [22] M . D. Adams and R. Ward. Two families of symmetry-preserving reversible integer-to-integer wavelet trans-forms. In Proc. o f IEEE International Symposium on Circuits and Systems, volume 2, pages 600-603, Scotts-dale, AZ, USA, May 2002. [23] R. Ansari, E. Ceran, and N. Memon. Near-lossless image compression techniques. In Proc. ofSPIE: Visual Communications and Image Processing, volume 3309, pages 731-742, January 1998. [24] M . Antonini, M . Barlaud, P. Mathieu, and I. Daubechies. Image coding using wavelet transform. IEEE Trans, on Image Processing, l(2):205-220, April 1992. [25] J. Askelof, M . Larsson Carlander, and C. Christopoulos. Region of interest coding in JPEG 2000. Signal Processing: Image Communication, 17(1): 105-111, January 2002. [26] E. Atsumi and N. Farvardin. Lossy/lossless region-of-interest image coding based on set partitioning in hierar-chical trees. In Proc. of IEEE International Conference on Image Processing, volume 1, pages 87-91, Chicago, IL, USA, October 1998. [27] B. Berthelot, E. Majani, and P. Onno. Transform core experiment: Compare the SNR performance of various reversible integer wavelet transforms. ISO/TEC JTC 1/SC 29/WG 1 N 902, June 1998. [28] H. Blume and A. Fand. Reversible and irreversible image data compression using the S-transform and Lempel-Ziv coding. In Proc. ofSPIE, volume 1091, pages 2-18, Newport Beach, CA, USA, January 1989. [29] M . Boliek, M . J. Gormish, E. L. Schwartz, and A. Keith. Next generation image compression and manipulation using CREW. In Proc. of IEEE International Conference on Image Processing, volume 3, pages 567-570, Santa Barbara, CA, USA, October 1997. [30] A. D. Booth. A signed binary multiplication technique. Quarterly Journal of Mechanics and Applied Mathe-matics, 4:236-240,1951. [31] C. Brislawn, S. Mniszewski, M . Pal, A. Percus, B. Wohlberg, T. Acharya, P.-S. Tsai, and M . Lepley. Even-length filter bank option. ISO/TEC JTC 1/SC 29/WG 1 N 2209, July 2001. [32] C. Brislawn and B. Wohlberg. Boundary extensions and reversible implementation for half-sample symmetric filter banks. ISO/TEC JTC 1/SC 29/WG 1 N 2119, March 2001. [33] C. M . Brislawn. Preservation of subband symmetry in multirate signal coding. IEEE Trans, on Signal Process-ing, 43(12):3046-3050, December 1995. [34] C. M . Brislawn. Classification of nonexpansive symmetric extension transforms for multirate filter banks. Applied and Computational Harmonic Analysis, 3(4):337-357, October 1996. Bibliography 112 [35] C. M . Brislawn, J. N . Bradley, R. J. Onyshczak, and T. Hopper. The FBI compression standard for digitized fingerprint images. Preprint, 1996. [36] F. A. M . L. Bruekers and A. W. M . van den Enden. New networks for perfect inversion and perfect reconstruc-tion. IEEE Journal on Selected Areas in Communications, 10(1):130—137, January 1992. [37] C. S. Burrus, R. A. Gopinath, and H. Guo. Introduction to Wavelets and Wavelet Transforms: A Primer. Prentice Hall, Upper Saddle River, NJ, USA, 1998. [38] A. R. Calderbank, I. Daubechies, W. Sweldens, and B.-L. Yeo. Wavelet transforms that map integers to integers. Technical report, Department of Mathematics, Princeton University, August 1996. [39] A. R. Calderbank, I. Daubechies, W. Sweldens, and B.-L. Yeo. Lossless image compression using integer to integer wavelet transforms. In Proc. o f IEEE International Conference on Image Processing, volume 1, pages 596-599, Santa Barbara, CA, USA, October 1997. [40] A. R. Calderbank, I. Daubechies, W. Sweldens, and B.-L. Yeo. Wavelet transforms that map integers to integers. Applied and Computational Harmonic Analysis, 5(3):332-369, July 1998. [41] B. Carpentieri, M . J. Weinberger, and G. Seroussi. Lossless compression of continuous-tone images. Proc. of IEEE, 88(11):1797-1809, November 2000. [42] J. J. F. Cavanagh. Digital Computer Arithmetic: Design and Implementation. McGraw-Hill, New York, 1984. [43] H. Cha and L. F. Chaparro. Adaptive morphological representation of signals: polynomial and wavelet methods. Multidimensional Systems and Signal Processing, 8(3):249-271, July 1997. [44] H. Chao, P. Fisher, and Z. Hua. An approach to integer wavelet transforms for lossless for image compression. In Proc. of International Symposium on Computational Mathematics, pages 19-38, Guangzhou, China, August 1997. [45] T. Chen and P. P. Vaidyanathan. Multidimensional multirate filters and filter banks derived from one dimensional filters. IEEE Trans, on Signal Processing, 41(5): 1749-1765, May 1993. [46] C. Christopoulos, J. Askelof, and M . Larsson. Efficient methods for encoding regions of interest in the upcoming JPEG 2000 still image coding standard. IEEE Signal Processing Letters, 7(9):247-249, September 2000. [47] C. Christopoulos, A. Skodras, and T. Ebrahimi. The JPEG 2000 still image coding system: An overview. IEEE Trans, on Consumer Electronics, 46(4): 1103-1127, November 2000. [48] T. Cooklev, A. Nishihara, T. Yoshida, and M . Sablatash. Multidimensional two-channel linear phase FIR filter banks and wavelet bases with vanishing moments. Multidimensional Systems and Signal Processing, 9(1):39-76, January 1998. [49] I. Daubechies and W. Sweldens. Factoring wavelet transforms into lifting steps. Journal of Fourier Analysis and Applications, 4:247-269,1998. [50] S. Dewitte and J. Cornells. Lossless integer wavelet transform. IEEE Signal Processing Letters, 4(6): 158-160, June 1997. [51] S. Dewitte, P. De Muynck, A. Munteanu, and J. Cornells. Lossless image subband coding with an integer wavelet transform. In Proc. of International Conference on Digital Signal Processing, volume 2, pages 555-557, Santorini, Greece, July 1997. [52] E. Dubois. The sampling and reconstruction of time-varying imagery with application in video systems. Proc. of IEEE, 73(4):502-522, April 1985. [53] D. E. Dudgeon and R. M . Mersereau. Multidimensional Digital Signal Processing. Prentice-Hall, Englewood Cliffs, NJ, USA, 1984. Bibliography 113 [54] O. Egger, W. L i , and M . Kunt. High compression image coding using an adaptive morphological subband decomposition. Proc. o f IEEE, 83(2):272-287, February 1995. [55] D. Le Gall and A. Tabatabai. Sub-band coding of digital images using symmetric short kernel filters and arithmetic coding techniques. In Proc. of IEEE International Conference on Acoustics, Speech, and Signal Processing, volume 2, pages 761-764, New York, NY, USA, April 1988. [56] R. S. Garfinkel and G. L. Nemhauser. Integer Programming. John Wiley and Sons, Toronto, ON, Canada, 1972. [57] O. N. Gerek and A. E. Cetin. Linear/nonlinear adaptive polyphase subband decomposition structures for im-age compression. In Proc. of IEEE International Conference on Acoustics, Speech, and Signal Processing, volume 3, pages 1345-1348, Seattle, WA, USA, May 1998. [58] O. N. Gerek and A. E. Cetin. Adaptive polyphase subband decomposition structures for image compression. IEEE Trans, on Image Processing, 9(10): 1649-1660, October 2000. [59] M . J. Gormish, E. L. Schwartz, A. F. Keith, M . R Boliek, and A. Zandi. Lossless and nearly lossless compression of high-quality images. In Proc. ofSPIE, volume 3025, pages 62-70, San Jose, CA, USA, March 1997. [60] J. Goutsias and H. J. A. M . Heijmans. Nonlinear multiresolution signal decomposition schemes—part i: Mor-phological pyramids. IEEE Trans, on Image Processing, 9(11): 1862-1876, November 2000. [61] Independent JPEG Group. JPEG library (version 6b). Available online from http: / /www. i jg . org, 2000. [62] A. Haar. Zur theorie der orthogonalen funktionen-systeme. Math. Annal., 69:331-371, 1910. [63] P. Hao and Q. Shi. Matrix factorizations for reversible integer mapping. IEEE Trans, on Signal Processing, 49( 10):2314-2324, October 2001. [64] R. He, J. Bai, and D. Ye. Properties for perfect reconstruction using multirate linear phase two-dimensional nonseparable filter banks and its implementation. Optical Engineering, 37(8):2363-2375, August 1998. [65] H. J. A. M . Heijmans and J. Goutsias. Nonlinear multiresolution signal decomposition schemes—part ii: Mor-phological wavelets. IEEE Trans, on Image Processing, 9(11): 1897—1913, November 2000. [66] J. S. Houchin and D. W. Singer. File format technology in JPEG 2000 enables flexible use of still and motion sequences. Signal Processing: Image Communication, 17(1): 131—144, January 2002. [67] ISO/TEC call for contributions, JPEG 2000. ISO/TEC JTC 1/SC 29/WG 1 N505, March 1997. [68] JPEG-2000 test images. ISO/IEC JTC 1/SC 29/WG 1 N 545, July 1997. [69] ISO/IEC 10918-1:1994, Information technology—Digital compression and coding o f continuous-tone still im-ages: Requirements and guidelines, 1994. [70] ISO/IEC 14492: Information technology—LossyAossless coding ofbi-level images, 2001. [71] ISO/IEC 14495-1: Lossless and near-lossless compression of continuous-tone still images: Baseline, 2000. [72] ISO/IEC 15444-1: Information technology—JPEG 2000 image coding system—Part 1: Core coding system, 2000. [73] ISO/IEC 15444-2: Information technology—JPEG 2000 image coding system—Part 2: Extensions, 2002. [74] ISO/IEC 15444-3: Information technology—JPEG 2000 image coding system—Part 3: Motion JPEG 2000, 2002. [75] ISO/IEC 15444-4: Information technology—JPEG 2000 image coding system—Part 4: Compliance testing, 2002. [76] ISO/IEC 15444-5: Information technology—JPEG 2000 image coding system—Part 5: Reference software, 2002. Bibliography 114 [77] N. S. Jayant and P. Noll. Digital Coding of Waveforms: Principals and Applications to Speech and Video. Prentice-Hall, Englewood Cliffs, NJ, USA, 1984. [78] H.-Y. Jung and R. Prost. Rounding transform based approach for lossless subband coding. In Proc. of IEEE International Conference on Image Processing, volume 2, pages 274-277, Santa Barbara, CA, USA, October 1997. [79] H.-Y. Jung and R. Prost. Lossless subband coding system based on rounding transform. IEEE Trans, on Signal Processing, 46(9):2535-2540, September 1998. [80] H.-Y. Jung and R. Prost. Multi-port filtering system for lossless image compression. In Proc. of IEEE Interna-tional Conference on Image Processing, volume 3, pages 861-865, Chicago, DL, USA, October 1998. [81] A. A. C. Kalker and I. A. Shah. Ladder structures for multidimensional linear phase perfect reconstruction filter banks and wavelets. In Proc. ofSPIE, volume 1818, pages 12-20, Boston, MA, USA, November 1992. [82] T. A. C. M . Kalker and I. A. Shah. On ladder structures and linear phase conditions for multidimensional bi-orthogonal filter banks. Preprint, June 1993. [83] G. Karlsson and M . Vetterli. Theory of two dimensional multirate filter banks. IEEE Trans, on Acoustics, Speech, and Signal Proceesing, 38:925-937, June 1990. [84] J. H. Kasner, M . W. Marcellin, and B. R. Hunt. Universal trellis coded quantization. IEEE Trans, on Image Processing, 8(12): 1677-1687, December 1999. [85] H. Kim and C. C. L i . A fast reversible wavelet image compressor. In Proc. of SPIE, volume 2825, pages 929-940, 1996. [86] H. Kim and C. C. L i . Lossless and lossy image compression using biorthogonal wavelet transforms with multiplierless operations. IEEE Trans, on Circuits and Systems, 45(8): 1113-1118, August 1998. [87] K. Komatsu and K. Sezaki. Reversible subband coding of images. In Proc. of SPIE, volume 2501, pages 676-684, 1995. [88] K. Komatsu and K. Sezaki. Design of lossless block transforms and filter banks for image coding. IEICE Trans. Fundamentals, E82-A(8): 1656-1664, August 1999. [89] K. Komatsu and K. Sezaki. Lossless filter banks based on two point transform and interpolative prediction. In Proc. of IEEE International Conference on Acoustics, Speech, and Signal Processing, volume 3, pages 1469-1472, Phoenix, AZ, USA, March 1999. [90] J. Kovacevic and W. Sweldens. Wavelet families of increasing order in arbitrary dimensions. IEEE Trans, on Image Processing, 9(3):480-496, March 2000. [91] J. Kovacevic and M . Vetterli. Nonseparable multidimensional perfect reconstruction filter banks and wavelet bases fori?". IEEE Trans, on Information Theory, 38(2):533-555, March 1992. [92] A. S. Lewis and G. Knowles. Image compression using the 2-D wavelet transform. IEEE Trans, on Image Processing, l(2):244-250, April 1992. [93] X. L i , B. Tao, and M . T. Orchard. On implementing transforms from integers to integers. In Proc. of IEEE International Conference on Image Processing, volume 3, pages 881-885, Chicago, IL, USA, October 1998. [94] J. S. Lim. Two-Dimensional Signal and Image Processing. Prentice Hall, Englewood Cliffs, NJ, USA, 1990. [95] Y.-P. Lin and P. P. Vaidyanathan. Theory and design of two-dimensional filter banks: A review. Multidimen-sional Systems and Signal Processing, 7(3-4):263-330, July 1996. [96] P. Lux. A novel set of closed orthogonal functions for picture coding. Archiv f u r Elektronik und Uebertragung-stechnik, 31(7):267-274, July 1977. Bibliography 115 [97] M . W. Marcellin, M . A. Lepley, A. Bilgin, T. J. Flohr, T. T. Chinen, and J. H. Kasner. An overview of quanziation in JPEG 2000. Signal Processing: Image Communication, 17(1):73—84, January 2002. [98] T. G. Marshall. Zero-phase filter bank and wavelet code r matrices: Properties, triangular decompositions, and a fast algorithm. Multidimensional Systems and Signal Processing, 8:71-88, 1997. [99] S. A. Martucci and R. M . Mersereau. The symmetric convolution approach to the nonexpansive implementation of FTR filter banks for images. In Proc. of IEEE International Conference on Acoustics, Speech, and Signal Processing, volume 5, pages 65-68, Minneapolis, M N , USA, April 1993. [100] N. Memon, X . Wu, and B. -L. Yeo. Improved techniques for lossless image compression with reversible integer wavelet transforms. In Proc. of IEEE International Conference on Image Processing, volume 3, pages 891-895, Chicago, IL, USA, October 1998. [101] A. Munteanu, J. Cornelis, P. De Muynck, and P. Cristea. Wavelet lossless compression of coronary angiographic images. In Computers in Cardiology, volume 24, pages 183-186, September 1997. [102] D. Nister and C. Christopoulos. Lossless region of interest with a naturally progressive still image coding algorithm. In Proc. of IEEE International Conference on Image Processing, volume 3, pages 856-860, Chicago, IL, USA, October 1998. [103] University of British Columbia. JPEG-LS codec software. Available online from http: / /spmg. ece. ubc. ca, 2000. version 2.2. [104] S. Oraintara, Y. J. Chen, and T. Q. Nguyen. Integer fast Fourier transform. IEEE Trans, on Signal Processing, 50(3):607-618, 2002. [105] H . Park, T. Kalker, and M . Vetterli. Grobner bases and multidimensional FIR multirate systems. Multidimen-sional Systems and Signal Processing, 8(1-2): 11-30, 1997. [106] H. Park and C. Woodburn. An algorithmic proof of Suslin's stability theorem for polynomial rings. Journal of Algebra, 178(l):277-298, November 1995. [107] W. B. Pennebaker and J. L . Mitchell. JPEG: Still Image Data Compression Standard. Kluwer Academic, New York, NY, USA, 1992. [108] M . Rabbani and P. W. Jones. Digital Image Compression Techniques. SPIE Optical Engineering Press, Belling-ham, W A . U S A , 1991. [109] M . Rabbani and R. Joshi. A n overview of the JPEG2000 still image compression standard. Signal Processing: Image Communication, 17(1):3—48, January 2002. [110] S. Ranganath and H . Blume. Hierarchical image decomposition and filtering using the s-transform (radiological images). In Proc. ofSPIE, volume 914, pages 799-814, Newport Beach, C A , USA, January 1988. [I l l ] D. W. Redmill and D. R. Bull. Multiplier-free linear phase filter banks for image and video compression. In Proc. ofSPIE, volume 3309, pages 634-644, San Jose, C A , USA, January 1998. [112] SAIC and University of Arizona. JPEG-2000 V M 0 software. ISO/IEC JTC 1/SC 29/WG 1 N 840, May 1998. [113] SAIC and University of Arizona. JPEG-2000 V M software (version 7.1). ISO/IEC JTC 1/SC 29/WG 1 N 1691, April 2000. [114] A. Said and W. A. Pearlman. An image multiresolution representation for lossless and lossy compression. IEEE Trans, on Image Processing, 5(9):1303—1310, September 1996. [115] A. Said and W. A. Pearlman. A new fast and efficient image codec based on set partitioning in hierarchical trees. IEEE Trans, on Circuits and Systems for Video Technology, 6(3):243-250, June 1996. [116] D. Santa-Cruz, R. Grosbois, and T. Ebrahimi. JPEG 2000 performance evaluation and assessment. Signal Processing: Image Communication, 17(1): 113-130, January 2002. Bibliography 116 [117] I. Shah and A. Kalker. Theory and design of multidimensional QMF sub-band filters from 1-D niters and poly-nomials using transforms. IEE Proceedings I: Communications, Speech and Vision, 140(1):67-71, February 1993. [118] L A . Shah, O. Akiwumi-Assani, and B. Johnson. A chip set for lossless image compression. IEEE Journal of Solid-State Circuits, 26(3):237-244, March 1991. [119] J. M . Shapiro. Embedded image coding using zerotrees of wavelet coefficients. IEEE Trans, on Signal Process-ing, 41(12):3445-3462, December 1993. [120] F. Sheng, A. Bilgin, P. J. Sementilli, and M . W. Marcellin. Lossy and lossless image compression using re-versible integer wavelet transforms. In Proc. of IEEE International Conference on Image Processing, volume 3, pages 876-880, Chicago, IL, USA, October 1998. [121] J. E. Shockley. Introduction to Number Theory. Holt, Rinehart, and Winston, Toronto, ON, Canada, 1967. [122] M . J. T. Smith and S. L. Eddins. Subband coding of images with octave band tree structures. In Proc. of IEEE International Conference on Acoustics, Speech, and Signal Processing, pages 1382-1385, Dallas, TX, USA, 1987. [123] M . J. T. Smith and S. L. Eddins. Analysis/synthesis techniques for subband image coding. IEEE Trans, on Acoustics, Speech, and Signal Proceesing, 38(8): 1446-1456, August 1990. [124] SPIHT home page. Available online from http: //www. c i p r . r p i . edu/research/SPIHT, 2000. [125] G. Strang and T. Nguyen. Wavelets and Filter Banks. Wellesley-Cambridge Press, Wellesley, MA, USA, 1996. [126] N. Strobel, S. K. Mitra, and B. S. Manjunath. Progressive-resolution transmission and lossless compression of color images for digital image libraries. In Proc. of International Conference on Digital Signal Processing, volume 1, pages 435-438, Santorini, Greece, July 1997. [127] N. Strobel, S. K. Mitra, and B. S. Manjunath. Reversible wavelet and spectral transforms for lossless com-pression of color images. In Proc. of IEEE International Conference on Image Processing, volume 3, pages 896-900, Chicago, IL, USA, October 1998. [128] A. Suslin. On the structure of the special linear group over polynomial rings. Mathematics of the USSR Izvestija, 11:221-238, 1977. [129] W. Sweldens. The lifting scheme: A new philosophy in biorthogonal wavelet constructions. In Proc. of SPIE, volume 2569, pages 68-79, San Diego, CA, USA, September 1995. [130] W. Sweldens. The lifting scheme: A custom-design construction of biorthogonal wavelets. Applied and Com-putational Harmonic Analysis, 3(2): 186-200, 1996. [131] W. Sweldens. The lifting scheme: A construction of second generation wavelets. SIAM Journal of Mathematical Analysis, 29(2):511-546, March 1998. [132] D. Taubman. High performance scalable image compression with EBCOT. In Proc. of IEEE International Conference on Image Processing, volume 3, pages 344-348, Kobe, Japan, 1999. [133] D. Taubman. High performance scalable image compression with EBCOT. IEEE Trans, on Image Processing, 9(7): 1158-1170, July 2000. [134] D. Taubman, E. Ordentlich, M . Weinberger, and G. Seroussi. Embedded block coding in JPEG 2000. Signal Processing: Image Communication, 17(l):49-72, January 2002. [135] D. S. Taubman and M . W. Marcellin. JPEG2000: Image Compression Fundamentals, Standards and Practice. Kluwer Academic, Boston, MA, USA, 2002. [136] D. B. H. Tay and N. G. Kingbury. Flexible design of multidimensional perfect reconstruction FIR two-band filters using transformation of variables. IEEE Trans, on Image Processing, 2(4):466-480, October 1993. Bibliography 117 [137] L. Tolhuizen, H. Hollmann, and T. A. C. M. Kalker. On the realizability of biorthogonal m-dimensional two-band filter banks. IEEE Trans, on Signal Processing, 43(3):640-648, March 1995. [138] P. P. Vaidyanathan. Fundamentals of multidimensional multirate digital signal processing. Sadhana, 15(3): 157— 176, November 1990. [139] P. P. Vaidyanathan. Multirate digital filters, filter banks, polyphase networks, and applications: a tutorial. Proc. of IEEE, 78(l):56-93, January 1990. [140] P. P. Vaidyanathan. Multirate Systems and Filter Banks. Prentice-Hall, Englewood Cliffs, NJ, USA, 1993. [141] M. Vetterli and J. Kovacevic. Wavelets and Subband Coding. Prentice-Hall, Englewood Cliffs, NJ, USA, 1995. [142] J. D. Villasenor, B. Belzer, and J. Liao. Wavelet filter evaluation for image compression. IEEE Trans, on Image Processing, 4(8): 1053-1060, August 1995. [143] E. Viscito and J. P. Allebach. The analysis and design of multidimensional FIR perfect reconstruction filter banks for arbitrary sampling lattices. IEEE Trans, on Circuits and Systems, 38(1):29-41, January 1991. [144] G. K. Wallace. The JPEG still picture compression standard. Communications of the ACM, 34(4):30-44, April 1991. [145] M. J. Weinberger, G. Seroussi, and G. Sapiro. The LOCO-I lossless image compression algorithm: Principles and standarization into JPEG-LS. IEEE Trans, on Image Processing, 9(8): 1309-1324, August 2000. [146] T. Wendler. Verfahren fur die hierarchische Codierung von Einzelbildern in medizinischen Bildinformation-ssystemen. Dr. Ing. thesis, Rheinisch-Westfalische Technische Hochschule Aachen, Germany, 1987. [147] I. H. Witten, R. M. Neal, and J. G. Cleary. Arithmetic coding for data compression. Communications of the ACM, 30(6):520-540, 1987. [148] A. Zandi, J. D. Allen, E. L. Schwartz, and M. Boliek. CREW: Compression with reversible embedded wavelets. In Proc. of IEEE Data Compression Conference, pages 212-221, Snowbird, UT, USA, March 1995. [149] W. Zeng, S. Daly, and S. Lei. An overview of the visual optimization tools in JPEG 2000. Signal Processing: Image Communication, 17(1):85—104, January 2002. [150] Y. Zeng, L. Cheng, G. Bi, and A. C. Kot. Integer DCTs and fast algorithms. IEEE Trans, on Signal Processing, 49(ll):2774-2782,2001. I am enough of an artist to draw freely upon my imagination. Imagination is more important than knowledge. Knowledge is limited. Imagination encircles the world. —Albert Einstein (1879-1955) 118 Appendix A JPEG 2000: An International Standard for Still Image Compression Many that live deserve death. And some that die deserve life. Can you give it to them? Then do not be too eager to deal out death in judgement. For even the very wise cannot see all ends. — J . Ft. Ft. Tolkein, The Fellowship of the Ring, 1954 Overview The JPEG-2000 image compression standard is introduced, and the JPEG-2000 Part-1 codec is described, analyzed, and evaluated. The JasPer software is also discussed briefly. A.1 Introduction Digital imagery is pervasive in our world today. Consequently, standards for the efficient representation and inter-change of digital images are essential. To date, some of the most successful still image compression standards have resulted from the ongoing work of the Joint Photographic Experts Group (JPEG). This group operates under the aus-pices of Joint Technical Committee 1, Subcommittee 29, Working Group 1 (JTC 1/SC 29/WG 1), a collaborative effort between the International Organization for Standardization (ISO) and International Telecommunication Union Standardization Sector (ITU-T). Both the JPEG [69,144,107] and JPEG-LS [71,145,41 ] standards were born from the work of the JPEG committee. For the last few years, the JPEG committee has been working towards the establishment of a new standard known as JPEG 2000 (i.e., ISO/IEC 15444). The fruits of these labors are now coming to bear, as JPEG-2000 Part 1 (i.e., ISO/EEC 15444-1 [72]) has recently been approved as a new international standard. In this appendix, we provide a detailed technical description of the JPEG-2000 Part-1 codec, in addition to a brief overview of the JPEG-2000 standard. (Some of the material in this appendix has also appeared in [4, 14].) This exposition is intended to serve as a reader-friendly starting point for those interested in learning about JPEG 2000. Although many details are included in our presentation, some details are necessarily omitted. The reader should, therefore, refer to the standard [72] before attempting an implementation. The JPEG-2000 codec realization in the JasPer software [2,14,3] may also serve as a practical guide for implementors. (See Section A.6 for more information about JasPer.) The reader may also find [135,109,47] to be useful sources of information on the JPEG-2000 standard. The remainder of this appendix is structured as follows. Section A.2 begins with a overview of the JPEG-2000 standard. This is followed, in Section A.3, by a detailed description of the JPEG-2000 Part-1 codec. In Section A.4, the codec is evaluated. The complexity of its various components is analyzed, and its performance is compared to codecs from other existing standards. Finally, we conclude this appendix with a brief summary of our results and some closing remarks in Section A.5. Throughout our presentation, a basic understanding of image coding is assumed. Appendix A . JPEG 2000: A n International Standard for Still Image Compression 119 Table A . l : Parts of the standard Part Description 1 Core coding system [72]. Specifies the core (or minimum functionality) codec for the JPEG-2000 family of standards. 2 Extensions [73]. Specifies additional functionalities that are useful in some applications but need not be supported by all codecs. 3 Motion JPEG-2000 [74]. Specifies extensions to JPEG-2000 for intraframe-style video com-pression. 4 Conformance testing [75]. Specifies the procedure to be employed for compliance testing. 5 Reference software [76]. Provides sample software implementations of the standard to serve as a guide for implementors. A.2 JPEG 2000 The JPEG-2000 standard supports lossy and lossless compression of single-component (e.g., grayscale) and multi-component (e.g., color) imagery. In addition to this basic compression functionality, however, numerous other features are provided, including: 1) progressive recovery of an image by fidelity or resolution; 2) region of interest coding, whereby different parts of an image can be coded with differing fidelity; 3) random access to particular regions of an image without needing to decode the entire code stream; 4) a flexible file format with provisions for specifying opacity information and image sequences; and 5) good error resilience. Due to its excellent coding performance and many attractive features, JPEG 2000 has a very large potential application base. Some possible application areas include: image archiving, Internet, web browsing, document imaging, digital photography, medical imaging, remote sensing, and desktop publishing. A.2.1 Why JPEG 2000? Work on the JPEG-2000 standard commenced with an initial call for contributions [67] in March 1997. The purpose of having a new standard was twofold. First, it would address a number of weaknesses in the existing JPEG standard. Second, it would provide a number of new features not available in the JPEG standard. The preceding points led to several key objectives for the new standard, namely that it should: 1) allow efficient lossy and lossless compression within a single unified coding framework, 2) provide superior image quality, both objectively and subjectively, at low bit rates, 3) support additional features such as region of interest coding, and a more flexible file format, 4) avoid excessive computational and memory complexity. Undoubtedly, much of the success of the original JPEG standard can be attributed to its royalty-free nature. Consequently, considerable effort has been made to ensure that a minimally-compliant JPEG-2000 codec can be implemented free of royalties1. A.2.2 Structure of the Standard The JPEG-2000 standard is comprised of numerous parts, several of which are listed in Table A . l . For convenience, we will refer to the codec defined in Part 1 of the standard as the baseline codec. The baseline codec is simply the core (or minimal functionality) coding system for the JPEG-2000 standard. Parts 2 and 3 describe extensions to the baseline codec that are useful for certain specific applications such as intraframe-style video compression. In this appendix, we will, for the most part, limit our discussion to the baseline codec. Some of the extensions proposed for inclusion in the JPEG-2000 Part-2 standard (i.e., [73]) will be discussed briefly. Unless otherwise indicated, our exposition considers only the baseline system. For the most part, the JPEG-2000 standard is written from the point of view of the decoder. That is, the decoder is defined quite precisely with many details being normative in nature (i.e., required for compliance), while many parts of the encoder are less rigidly specified. Obviously, implementors must make a very clear distinction between normative and informative clauses in the standard. For the purposes of our discussion, however, we will only make such distinctions when absolutely necessary. 'Whether these efforts ultimately prove successful remains to be seen, however, as there are still some unresolved intellectual property issues at the time of this writing. Appendix A. JPEG 2000: An International Standard for Still Image Compression 120 Component i (b) (a) Figure A . l : Source image model, (a) An image with N components, (b) Individual component. A.3 JPEG-2000 Codec Having briefly introduced the JPEG-2000 standard, we are now in a position to begin examining the JPEG-2000 codec in detail. The codec is based on wavelet/subband coding techniques [92,24]. It handles both lossy and lossless compression using the same transform-based framework, and borrows heavily on ideas from the embedded block coding with optimized truncation (EBCOT) scheme [132,133, 134]. In order to facilitate both lossy and lossless coding in an efficient manner, reversible ITI [40,59] and nonreversible real-to-real transforms are employed. To code transform data, the codec makes use of bit-plane coding techniques. For entropy coding, a context-based adaptive binary arithmetic coder [ 1 4 7 ] is used—more specifically, the M Q coder from the JBIG2 standard (i.e., ISO/IEC 14492 [70]). Two levels of syntax are employed to represent the coded image: a code stream and file format syntax. The code stream syntax is similar in spirit to that used in the JPEG standard. The remainder of Section A.3 is structured as follows. First, Sections A.3.1 to A.3.3, discuss the source image model and how an image is internally represented by the codec. Next, Section A.3.4 examines the basic structure of the codec. This is followed, in Sections A.3.5 to A.3.13 by a detailed explanation of the coding engine itself. Next, Sections A.3.14 and A.3.15 explain the syntax used to represent a coded image. Finally, Section A.3.16 briefly describes some extensions proposed for inclusion in the JPEG-2000 Part-2 standard. A.3.1 Source Image Model Before examining the internals of the codec, it is important to understand the image model that it employs. From the codec's point of view, an image is comprised of one or more components (up to a limit of 2 1 4 ) , as shown in Figure A . 1(a). As illustrated in Figure A . 1(b), each component consists of a rectangular array of samples. The sample values for each component are integer valued, and can be either signed or unsigned with a precision from 1 to 38 bits/sample. The signedness and precision of the sample data are specified on a per-component basis. A l l of the components are associated with the same spatial extent in the source image, but represent different spectral or auxiliary information. For example, a RGB color image has three components with one component rep-resenting each of the red, green, and blue color planes. In the simple case of a grayscale image, there is only one component, corresponding to the luminance plane. The various components of an image need not be sampled at the same resolution. Consequently, the components themselves can have different sizes. For example, when color images are represented in a luminance-chrominance color space, the luminance information is often more finely sampled than the chrominance data. A.3.2 Reference Grid Given an image, the codec describes the geometry of the various components in terms of a rectangular grid called the reference grid. The reference grid has the general form shown in Figure A.2. The grid is of size Xsiz x Ysiz with the origin located at its top-left corner. The region with its top-left corner at (XOsiz, YOsiz) and bottom-right corner at (Xsiz - 1, Ysiz — 1) is called the image area, and corresponds to the picture data to be represented. The width and height of the reference grid cannot exceed 2 3 2 — 1 units, imposing an upper bound on the size of an image that can be handled by the codec. Appendix A. JPEG 2000: An International Standard for Still Image Compression 121 (0.0) (xosioro«iz) Image Area (Xsiz-l,Yiiz-l) Figure A.2: Reference grid. All of the components are mapped onto the image area of the reference grid. Since components need not be sampled at the full resolution of the reference grid, additional information is required in order to establish this mapping. For each component, we indicate the horizontal and vertical sampling period in units of the reference grid, denoted as XRsiz and YRsiz, respectively. These two parameters uniquely specify a (rectangular) sampling grid consisting of all points whose horizontal and vertical positions are integer multiples of XRsiz and YRsiz, respectively. All such points that fall within the image area, constitute samples of the component in question. Thus, in terms of its own coordinate system, a component will have the size ([jg^] - f$Sfsl) x ( T Y R S E ! _ T Y I S I Z I ) a n c * i t s top-left sample will correspond to the point ([^gfjf], [ Y R S I Z D • ^ o t e t n a t t n e r e f e r e n c e grid also imposes a particular alignment of samples from the various components relative to one another. From the diagram, the size of the image area is (Xsiz —XOsiz) x (Ysiz — YOsiz). For a given image, many combinations of the Xsiz, Ysiz, XOsiz, and YOsiz parameters can be chosen to obtain an image area with the same size. Thus, one might wonder why the XOsiz and YOsiz parameters are not fixed at zero while the Xsiz and Ysiz parameters are set to the size of the image. As it turns out, there are subtle implications to changing the XOsiz and YOsiz parameters (while keeping the size of the image area constant). Such changes affect codec behavior in several important ways, as will be described later. This behavior allows a number of basic operations to be performed efficiently on coded images, such as cropping, horizontal/vertical flipping, and rotation by an integer multiple of 90 degrees. A.3.3 Tiling In some situations, an image may be quite large in comparison to the amount of memory available to the codec. Consequently, it is not always feasible to code the entire image as a single atomic unit. To solve this problem, the codec allows an image to be broken into smaller pieces, each of which is independently coded. More specifically, an image is partitioned into one or more disjoint rectangular regions called tiles. As shown in Figure A.3, this partitioning is performed with respect to the reference grid by overlaying the reference grid with a rectangular tiling grid having horizontal and vertical spacings of XTsiz and YTsiz, respectively. The origin of the tiling grid is aligned with the point (XTOsiz, YTOsiz). Tiles have a nominal size of XTsiz x YTsiz, but those bordering on the edges of the image area may have a size which differs from the nominal size. The tiles are numbered in raster scan order (starting at zero). By mapping the position of each tile from the reference grid to the coordinate systems of the individual compo-nents, a partitioning of the components themselves is obtained. For example, suppose that a tile has an upper left corner and lower right corner with coordinates (txo,ty0) and (txi — 1, tyx — 1), respectively. Then, in the coordinate space of a particular component, the tile would have an upper left corner and lower right corner with coordinates (tcxo,tcy0) and (tcxi — l.tcy, — 1), respectively, where (tcxo,tcy0) = (ftx0/XRsiz],[ty0/YRsizl) and (A.la) (tcxi.tcy,) = ([txi/XRsiz], [ty,/YRsiz]). (A.lb) These equations correspond to the illustration in Figure A.4. The portion of a component that corresponds to a single tile is referred to as a tile-component. Although the tiling grid is regular with respect to the reference grid, it is important to note that the grid may not necessarily be regular with respect to the coordinate systems of the components. Appendix A. JPEG 2000: An International Standard for Still Image Compression Xsiz Ysiz (0,0) (XTOsiz.YTOsiz) <XOsiz,YO To it) T 3 T 5 \ T, YTsiz XTOsiz XTsiz XTsiz Figure A.3: Tiling on the reference grid. (0,0) I " ' • <tcx0.tcy0 ) Tite-Component Data (tcxj-l.tey, Figure A.4: Tile-component coordinate system. Appendix A . JPEG 2000: An International Standard for Still Image Compression 123 Original Image Preprocessing Forward Intercomponent Transform Forward Intracomponent Transform Rate Control Quantization Tier-1 Encoder Tier-2 Encoder Coded Image (a) Coded Image Tier-2 Decoder Tier-1 Decoder Dequantization Inverse Intracomponent Transform Inverse Intercomponent Transform Postprocessing Reconstructed Image (b) Figure A.5: Codec structure. The structure of the (a) encoder and (b) decoder. A.3.4 Codec Structure The general structure of the codec is shown in Figure A.5 with the form of the encoder given by Figure A.5(a) and the decoder given by Figure A.5(b). From these diagrams, the key processes associated with the codec can be identified: 1) preprocessing/postprocessing, 2) intercomponent transform, 3) intracomponent transform, 4) quantiza-tion/dequantization, 5) tier-1 coding, 6) tier-2 coding, and 7) rate control. The decoder structure essentially mirrors that of the encoder. That is, with the exception of rate control, there is a one-to-one correspondence between functional blocks in the encoder and decoder. Each functional block in the decoder either exactly or approximately inverts the effects of its corresponding block in the encoder. Since tiles are coded independently of one another, the input image is (conceptually, at least) processed one tile at a time. In the sections that follow, each of the above processes is examined in more detail. A.3.5 Preprocessing/Postprocessing The codec expects its input sample data to have a nominal dynamic range that is approximately centered about zero. The preprocessing stage of the encoder simply ensures that this expectation is met. Suppose that a particular com-ponent has P bits/sample. The samples may be either signed or unsigned, leading to a nominal dynamic range of [ -2 P ~ 1 ,2 P _ 1 — 1] or [0,2P — 1], respectively. If the sample values are unsigned, the nominal dynamic range is clearly not centered about zero. Thus, the nominal dynamic range of the samples is adjusted by subtracting a bias of 2p~l from each of the sample values. If the sample values for a component are signed, the nominal dynamic range is already centered about zero, and no processing is required. By ensuring that the nominal dynamic range is centered about zero, a number of simplifying assumptions could be made in the design of the codec (e.g., with respect to context modeling, numerical overflow, etc.). The postprocessing stage of the decoder essentially undoes the effects of preprocessing in the encoder. If the sample values for a component are unsigned, the original nominal dynamic range is restored. Lastly, in the case of lossy coding, clipping is performed to ensure that the sample values do not exceed the allowable range. A.3.6 Intercomponent Transform In the encoder, the preprocessing stage is followed by the forward intercomponent transform stage. Here, an inter-component transform can be applied to the tile-component data. Such a transform operates on all of the components together, and serves to reduce the correlation between components, leading to improved coding efficiency. Only two intercomponent transforms are defined in the baseline JPEG-2000 codec: the irreversible color transform (ICT) and reversible color transform (RCT). The ICT is nonreversible and real-to-real in nature, while the RCT is reversible and ITI. Both of these transforms essentially map image data from the RGB to YCrCb color space. The transforms are defined to operate on the first three components of an image, with the assumption that components 0, Appendix A . JPEG 2000: An International Standard for Still Image Compression 124 1, and 2 correspond to the red, green, and blue color planes. Due to the nature of these transforms, the components on which they operate must be sampled at the same resolution (i.e., have the same size). As a consequence of the above facts, the ICT and RCT can only be employed when the image being coded has at least three components, and the first three components are sampled at the same resolution. The ICT may only be used in the case of lossy coding, while the RCT can be used in either the lossy or lossless case. Even if a transform can be legally employed, it is not necessary to do so. That is, the decision to use a multicomponent transform is left at the discretion of the encoder. After the intercomponent transform stage in the encoder, data from each component is treated independently. The ICT is nothing more than the classic RGB to YCrCb color space transform. The forward transform is defined as [V0(x,y)~] T 0.299 0.587 0.114 ] [U0(x,y)'] (A.2) where Uo(x,y), U\ (x,y), and U2{x,y) are the input components corresponding to the red, green, and blue color planes, respectively, and Vo(x,y), V\ {x,y), and V2(x,y) are the output components corresponding to the Y, Cr, and Cb planes, respectively. The inverse transform can be shown to be (A.3) The RCT is simply a reversible ITI approximation to the ICT (similar to that proposed in [59]). The forward transform is given by ~v0{x,y)' 0.29') 0.587 0.114 'U0(x,y) Vi(x,y) -0.16875 -0.33126 0.5 Ui(x,y) V2(x,y) 0.5 -0.41869 -0.08131 U2{x,y)_ ~U0(x,y)' "1 0 1.402 'Vo(x,y)~ Ui(x,y) = 1 -0.34413 -0.71414 Vi(x,y) U2(x,y)m 1 -1.772 0 Vi(x,y) Vo(x,y) = [l(U0(x,y)+2Ui(x,y)+U2(x,y))\, Vi(x,y) = U2(x,y)-Ui(x,y), and V2(x,y) = U0{x,y) - Ui(x,y), (A.4a) (A.4b) (A.4c) where Uo(x,y), Ui(x,y), l)2{x,y), Vo{x,y), V*i(jc,y), and V2(x,y) are defined as above. The inverse transform can be shown to be Ux{x,y) = V0(x,y)- [±(Vi{x,y) + V2(x,y))\, Uo(x,y) = V2{x,y) + U\{x,y), and U2(x,y) = Vl(x,y) + Ul(x,y). (A.5a) (A.5b) (A.5c) The inverse intercomponent transform stage in the decoder essentially undoes the effects of the forward intercom-ponent transform stage in the encoder. If a multicomponent transform was applied during encoding, its inverse is applied here. Unless the transform is reversible, however, the inversion may only be approximate due to the effects of finite-precision arithmetic. A.3.7 Intracomponent Transform Following the intercomponent transform stage in the encoder is the intracomponent transform stage. In this stage, transforms that operate on individual components can be applied. The particular type of operator employed for this purpose is the wavelet transform. Through the application of the wavelet transform, a component is split into numerous frequency bands (i.e., subbands). Due to the statistical properties of these subband signals, the transformed data can usually be coded more efficiently than the original untransformed data. Both reversible ITI [40,44,1,16] and nonreversible real-to-real wavelet transforms are employed by the baseline codec. The basic building block for such transforms is the one-dimensional (ID) two-channel perfect-reconstruction (PR) uniformly maximally-decimated (UMD) filter bank which has the general form shown in Figure A.6. Here, we focus on the lifting realization of the UMD filter bank [130,131], as it can be used to implement the reversible ITI and nonreversible real-to-real wavelet transforms employed by the baseline codec. In fact, for this reason, it is likely that this realization strategy will be employed by many codec implementations. The analysis side of the UMD filter bank, depicted in Figure A.6(a), is associated with the forward transform, while the synthesis side, depicted in Figure A.6(b), is associated with the inverse transform. In the diagram, the {Ai(z)}^~01, {Qi(x)}^, and {*,},L0 Appendix A . JPEG 2000: An International Standard for Still Image Compression 125 12 Mz) Qi K_2(z) „ I—I yo" Qx-Qo 12 yo[n\ Mz) Qx-2 Qx-i Ml(z) AxMz) Qx-2 (a) Mz) Ao(z) Qo T2 (b) Figure A.6: Lifting realization of ID two-channel UMD filter bank, (a) Analysis side, (b) Synthesis side. denote filter transfer functions, quantization operators, and (scalar) gains, respectively. To obtain ITI mappings, the {Qi(x)}}~01 are selected such that they always yield integer values, and the {i i}; = 0 are chosen as integers. For real-to-real mappings, the {Qi(x)}f~d are simply chosen as the identity, and the {j/},Lo are selected from the real numbers. To facilitate filtering at signal boundaries, symmetric extension [33,34] is employed. Since an image is a two-dimensional (2D) signal, clearly we need a 2D UMD filter bank. By applying the ID UMD filter bank in both the horizontal and vertical directions, a two-dimensional UMD filter bank is effectively obtained. The wavelet transform is then calculated by recursively applying the 2D UMD filter bank to the lowpass subband signal obtained at each level in the decomposition. Suppose that a (R— l)-level wavelet transform is to be employed. To compute the forward transform, we apply the analysis side of the 2D UMD filter bank to the tile-component data in an iterative manner, resulting in a number of subband signals being produced. Each application of the analysis side of the 2D UMD filter bank yields four subbands: 1) horizontally and vertically lowpass (LL), 2) horizontally lowpass and vertically highpass (LH), 3) horizontally highpass and vertically lowpass (HL), and 4) horizontally and vertically highpass (HH). A (R— l)-level wavelet decomposition is associated with R resolution levels, numbered from 0 to R— 1, with 0 and R— 1 corresponding to the coarsest and finest resolutions, respectively. Each subband of the decomposition is identified by its orientation (e.g., LL, LH, HL, HH) and its corresponding resolution level (e.g., 0,1,...,R— 1). The input tile-component signal is considered to be the LL«_i band. At each resolution level (except the lowest) the L L band is further decomposed. For example, the LL«_i band is decomposed to yield the LL/j_2, L H «_2 , HL/e_2, and HH/f_2 bands. Then, at the next level, the L L « _ 2 band is decomposed, and so on. This process repeats until the L L 0 band is obtained, and results in the subband structure illustrated in Figure A.7. In the degenerate case where no transform is applied, R = 1, and we effectively have only one subband (i.e., the LLo band). As described above, the wavelet decomposition can be associated with data at R different resolutions. Suppose that the top-left and bottom-right samples of a tile-component have coordinates (tcxo,tcy0) and (tcxi - l,tcy, — 1), respectively. This being the case, the top-left and bottom-right samples of the tile-component at resolution r have coordinates (trxo,try0) and (trxi - ljtry, — 1), respectively, given by (trxo,try0) = ([tcxo/2*-'- 1], [tcyo/2*"'-1]) and (tnctry,) = ( [ t c x , / ^ — 1 ] , [tcy 1/2*-'- 1l) (A.6a) (A.6b) where r is the particular resolution of interest. Thus, the tile-component signal at a particular resolution has the size (trxi -trxo) x (try, - t ry 0 ) . Not only are the coordinate systems of the resolution levels important, but so too are the coordinate systems for the various subbands. Suppose that we denote the coordinates of the upper left and lower right samples in a subband Appendix A . JPEG 2000: An International Standard for Still Image Compression 126 L L 0 HL R _ 2 HL R _, L H R - 2 HH R _ 2 L H R _ , H H R . , Figure A.7: Subband structure. as (tbx0,tby0) and (tbxi - l,tby, - 1), respectively. These quantities are computed as (tbxo,tby0)= { (tbxi,tby1) = < tcxp 1 r tcy0 1 \ 2 « - r - l i 2 " _ r _ 1 / tcxp _ 1~\ [ tcy0 2«- r -> 2 | ' I 2 R - r - ' tcxp 1 f t c y o texo _ i l f tcyQ 2 « - f - l 2 I ' | tcx, 1 f tcy, "1 \ I ' I tcxi _ \~\ [" tcyi 2R-r-1 2 ' 2* r 1 tcxi 1 [ tcy, 2«-r-1 ' 2 R r 1 for L L band for HL band for LH band ) for HH band, for L L band for HL band for LH band and (A.7a) (A.7b) ^ - i j . ^ - i j ) forHHband, where r is the resolution level to which the band belongs, R is the number of resolution levels, and tcxrj, tcy0, tcxi, and tcyt are as defined in (A.1). Thus, a particular band has the size (tbxj — tbxrj) x (tbyj — tby0). From the above equations, we can also see that (tbxo,tby0) = (trxo,try0) and (tbx^tby,) = (trxi,try,) for the LLr band, as one would expect. (This should be the case since the LLr band is equivalent to a reduced resolution version of the original data.) As will be seen, the coordinate systems for the various resolutions and subbands of a tile-component play an important role in codec behavior. By examining (A.1), (A.6), and (A.7), we observe that the coordinates of the top-left sample for a particular subband, denoted (tbxo,tby0), are partially determined by the XOsiz and YOsiz parameters of the reference grid. At each level of the decomposition, the parity (i.e., oddness/evenness) of tbxo and tby0 affects the outcome of the downsampling process (since downsampling is shift variant). In this way, the XOsiz and YOsiz parameters have a subtle, yet important, effect on the transform calculation. Having described the general transform framework, we now describe the two specific wavelet transforms supported by the baseline codec: the 5/3 and 9/7 transforms. The 5/3 transform is reversible, ITI, and nonlinear. This transform was proposed in [40], and is simply an approximation to a linear wavelet transform proposed in [55]. The 5/3 transform has an underlying ID UMD filter bank with the parameters: X = 2, A0(z) = -$ (z+l ) , A1(z) = \(l+z~l), Qo{x) = - [-x\, Qi (x) = [x+ j \ , and s0 = si = 1. (A.S) The 9/7 transform is nonreversible and real-to-real. This transform, proposed in [24], is also employed in the FBI fingerprint compression standard [35] (although the normalizations differ). The 9/7 transform has an underlying ID Appendix A . JPEG 2000: An International Standard for Still Image Compression 127 UMD filter bank with the parameters: X = 4, A0(z) = ao(z+1), i4i(z)=ai(l + z - , ) > (A.9) A2(z) = a 2 ( z + 1 ) , A3(z) = a 3(l+z"- 1), G,-(JC) = x fori = ' 0 , 1 , 2 , 3 , oto « - 1 . 5 8 6 1 3 4 , a] w - 0 . 0 5 2 9 8 0 , a 2 « 0 . 8 8 2 9 1 1 , a 3 w 0 . 4 4 3 5 0 6 , s0 « 1 .230174 , and J I = l/s0. Since the 5 /3 transform is reversible, it can be employed for either lossy or lossless coding. The 9 /7 transform, lacking the reversible property, can only be used for lossy coding. The number of resolution levels is a parameter of each transform. A typical value for this parameter is six (for a sufficiently large image). The encoder may transform all, none, or a subset of the components. This choice is at the encoder's discretion. The inverse intracomponent transform stage in the decoder essentially undoes the effects of the forward intracom-ponent transform stage in the encoder. If a transform was applied to a particular component during encoding, the corresponding inverse transform is applied here. Due to the effects of finite-precision arithmetic, the inversion process is not guaranteed to be exact unless reversible transforms are employed. A.3.8 Quantization/Dequantization In the encoder, after the tile-component data has been transformed (by intercomponent and/or intracomponent trans-forms), the resulting coefficients are quantized. Quantization allows greater compression to be achieved, by repre-senting transform coefficients with only the minimal precision required to obtain the desired level of image quality. Quantization of transform coefficients is one of the two primary sources of information loss in the coding path (the other source being the discarding of coding pass data as will be described later). Transform coefficients are quantized using scalar quantization with a deadzone. A different quantizer is employed for the coefficients of each subband, and each quantizer has only one parameter, its step size. Mathematically, the quantization process is defined as V(x,y) = [\U(x,y)\/A\ sgn£/(x,y) (A.10) where A is the quantizer step size, U(x,y) is the input subband signal, and V(x,y) denotes the output quantizer indices for the subband. Since this equation is specified in an informative clause of the standard, encoders need not use this precise formula. This said, however, it is likely that many encoders will, in fact, use the above equation. The baseline codec has two distinct modes of operation, referred to herein as integer mode and real mode. In integer mode, all transforms employed are ITI in nature (e.g., RCT, 5 /3 ) . In real mode, real-to-real transforms are employed (e.g., ICT, 9 /7 ) . In integer mode, the quantizer step sizes are always fixed at one, effectively bypassing quantization and forcing the quantizer indices and transform coefficients to be one and the same. In this case, lossy coding is still possible, but rate control is achieved by another mechanism (to be discussed later). In the case of real mode (which implies lossy coding), the quantizer step sizes are chosen in conjunction with rate control. Numerous strategies are possible for the selection of the quantizer step sizes, as will be discussed later in Section A.3.12. As one might expect, the quantizer step sizes used by the encoder are conveyed to the decoder via the code stream. In passing, we note that the step sizes specified in the code stream are relative and not absolute quantities. That is, the quantizer step size for each band is specified relative to the nominal dynamic range of the subband signal. In the decoder, the dequantization stage tries to undo the effects of quantization. Unless all of the quantizer step sizes are less than or equal to one, the quantization process will normally result in some information loss, and this inversion process is only approximate. The quantized transform coefficient values are obtained from the quantizer indices. Mathematically, the dequantization process is defined as U(x,y) = (V(x,y) + r s g n V ( x , y ) ) A (A.l l) where A is the quantizer step size, r is a bias parameter, V(x,y) are the input quantizer indices for the subband, and U(x,y) is the reconstructed subband signal. Although the value of r is not normatively specified in the standard, it is likely that many decoders will use the value of one half. Appendix A . JPEG 2000: A n International Standard for Still Image Compression 128 (0,0) ' «xcb' _y< ( m 2 , n 2 b ) • ( t b x 0 , tb; B o B, B2 B3 B4 B5 J l B7 ( t b x , - l B8 , t b y , - l ) • i i ycb' ycb' xcb' ~"cb' ^ X c b ' Figure A.8: Partitioning of a subband into code blocks. A.3.9 Tier-1 Coding After quantization is performed in the encoder, tier-1 coding takes place. This is the first of two coding stages. The quantizer indices for each subband are partitioned into code blocks. Code blocks are rectangular in shape, and their nominal size is a free parameter of the coding process, subject to certain constraints, most notably: 1) the nominal width and height of a code block must be an integer power of two, and 2) the product of the nominal width and height cannot exceed 4096. Suppose that the nominal code block size is tentatively chosen to be 2 x c b x 2 y c b . In tier-2 coding, yet to be discussed, code blocks are grouped into what are called precincts. Since code blocks are not permitted to cross precinct boundaries, a reduction in the nominal code block size may be required if the precinct size is sufficiently small. Suppose that the nominal code block size after any such adjustment is 2 x c b ' x 2 y c b ' where xcb' < xcb and ycb' < ycb. The subband is partitioned into code blocks by overlaying the subband with a rectangular grid having horizontal and vertical spacings of 2 x c b and 2 y c b , respectively, as shown in Figure A.8. The origin of this grid is anchored at (0,0) in the coordinate system of the subband. A typical choice for the nominal code block size is 64 x 64 (i.e., xcb = 6 and ycb = 6). Let us, again, denote the coordinates of the top-left sample in a subband as (tbxn,tby0). As explained in Sec-tion A.3.7, the quantity (tbxo, tby0) is partially determined by the reference grid parameters XOsiz and YOsiz. In turn, the quantity (tbxo,tby0) affects the position of code block boundaries within a subband. In this way, the XOsiz and YOsiz parameters have an important effect on the behavior of the tier-1 coding process (i.e., they affect the location of code block boundaries). After a subband has been partitioned into code blocks, each of the code blocks is independently coded. The coding is performed using the bit-plane coder described later in Section A.3.10. For each code block, an embedded code is produced, comprised of numerous coding passes. The output of the tier-1 encoding process is, therefore, a collection of coding passes for the various code blocks. On the decoder side, the bit-plane coding passes for the various code blocks are input to the tier-1 decoder, these passes are decoded, and the resulting data is assembled into subbands. In this way, we obtain the reconstructed quantizer indices for each subband. In the case of lossy coding, the reconstructed quantizer indices may only be approximations to the quantizer indices originally available at the encoder. This is attributable to the fact that the code stream may only include a subset of the coding passes generated by the tier-1 encoding process. In the lossless case, the reconstructed quantizer indices must be same as the original indices on the encoder side, since all coding passes must be included for lossless coding. Appendix A . JPEG 2000: An International Standard for Still Image Compression 129 v o do v o d , X h , h o x h , v l d 2 v i d , (a) (b) Figure A.9: Templates for context selection. The (a) 4-connected and (b) 8-connected neighbors. A.3.10 Bit-Plane Coding The tier-1 coding process is essentially one of bit-plane coding. After all of the subbands have been partitioned into code blocks, each of the resulting code blocks is independently coded using a bit-plane coder. Although the bit-plane coding technique employed is similar to those used in the embedded zerotree wavelet (EZW) [ 1 1 9 ] and set partitioning in hierarchical trees (SPIHT) [ 1 1 5 ] codecs, there are two notable differences: 1) no interband dependencies are exploited, and 2) there are three coding passes per bit plane instead of two. The first difference follows from the fact that each code block is completely contained within a single subband, and code blocks are coded independently of one another. By not exploiting interband dependencies, improved error resilience can be achieved. The second difference is arguably less fundamental. Using three passes per bit plane instead of two reduces the amount of data associated with each coding pass, facilitating finer control over rate. Also, using an additional pass per bit plane allows better prioritization of important data, leading to improved coding efficiency. As noted above, there are three coding passes per bit plane. In order, these passes are as follows: 1) significance, 2) refinement, and 3) cleanup. All three types of coding passes scan the samples of a code block in the same fixed order shown in Figure A. 10. The code block is partitioned into horizontal stripes, each having a nominal height of four samples. If the code block height is not a multiple of four, the height of the bottom stripe will be less than this nominal value. As shown in the diagram, the stripes are scanned from top to bottom. Within a stripe, columns are scanned from left to right. Within a column, samples are scanned from top to bottom. The bit-plane encoding process generates a sequence of symbols for each coding pass. Some or all of these symbols may be entropy coded. For the purposes of entropy coding, a context-based adaptive binary arithmetic coder is used—more specifically, the MQ coder from the JBIG2 standard (i.e., ISO/IEC 14492 [70]). For each pass, all of the symbols are either arithmetically coded or raw coded (i.e., the binary symbols are emitted as raw bits with simple bit stuffing). The arithmetic and raw coding processes both ensure that certain bit patterns never occur in the output, allowing such patterns to be used for error resilience purposes. Cleanup passes always employ arithmetic coding. In the case of the significance and refinement passes, two possibilities exist, depending on whether the so called arithmetic-coding bypass mode (also known as lazy mode) is enabled. If lazy mode is enabled, only the significance and refinement passes for the four most significant bit planes use arithmetic coding, while the remaining such passes are raw coded. Otherwise, all significance and refinement passes are arithmetically coded. The lazy mode allows the computational complexity of bit-plane coding to be significantly reduced, by decreasing the number of symbols that must be arithmetically coded. This comes, of course, at the cost of reduced coding efficiency. As indicated above, coding pass data can be encoded using one of two schemes (i.e., arithmetic or raw coding). Consecutive coding passes that employ the same encoding scheme constitute what is known as a segment. All of the coding passes in a segment can collectively form a single codeword or each coding pass can form a separate codeword. Which of these is the case is determined by the termination mode in effect. Two termination modes are supported: per-pass termination and per-segment termination. In the first case, only the last coding pass of a segment is terminated. In the second case, all coding passes are terminated. Terminating all coding passes facilitates improved error resilience at the expense of decreased coding efficiency. Since context-based arithmetic coding is employed, a means for context selection is necessary. Generally speaking, context selection is performed by examining state information for the 4-connected or 8-connected neighbors of a sample as shown in Figure A.9. In our explanation of the coding passes that follows, we focus on the encoder side as this facilitates easier under-standing. The decoder algorithms follow directly from those employed on the encoder side. Appendix A . JPEG 2000: An International Standard for Still Image Compression 130 1 for each sample in code block do 2 if sample previously insignificant and predicted to become significant during current bit plane then 3 code significance of sample /* 1 binary symbol */ 4 if sample significant then 5 code sign of sample /* 1 binary symbol */ 6 endif 7 endif 8 endfor Algorithm A . l : Significance pass algorithm. 1 for each sample in code block do 2 if sample found significant in previous bit plane then 3 code next most significant bit in sample /* 1 binary symbol */ 4 endif 5 endfor Algorithm A.2: Refinement pass algorithm. A.3.10.1 Significance Pass The first coding pass for each bit plane is the significance pass. This pass is used to convey significance and (as necessary) sign information for samples that have not yet been found to be significant and are predicted to become significant during the processing of the current bit plane. The samples in the code block are scanned in the order shown previously in Figure A. 10. If a sample has not yet been found to be significant, and is predicted to become significant, the significance of the sample is coded with a single binary symbol. If the sample also happens to be significant, its sign is coded using a single binary symbol. In pseudocode form, the significance pass is described by Algorithm A . l . If the most significant bit plane is being processed, all samples are predicted to remain insignificant. Otherwise, a sample is predicted to become significant if any 8-connected neighbor has already been found to be significant. As a consequence of this prediction policy, the significance and refinement passes for the most significant bit plane are always empty (and need not be explicitly coded). The symbols generated during the significance pass may or may not be arithmetically coded. If arithmetic coding is employed, the binary symbol conveying significance information is coded using one of nine contexts. The particular context used is selected based on the significance of the sample's 8-connected neighbors and the orientation of the subband with which the sample is associated (e.g., LL, LH, HL, HH). In the case that arithmetic coding is used, the sign of a sample is coded as the difference between the actual and predicted sign. Otherwise, the sign is coded directly. Sign prediction is performed using the significance and sign information for 4-connected neighbors. A.3.10.2 Refinement Pass The second coding pass for each bit plane is the refinement pass. This pass signals subsequent bits after the most significant bit for each sample. The samples of the code block are scanned using the order shown earlier in Figure A. 10. If a sample was found to be significant in a previous bit plane, the next most significant bit of that sample is conveyed using a single binary symbol. This process is described in pseudocode form by Algorithm A.2. Like the significance pass, the symbols of the refinement pass may or may not be arithmetically coded. If arithmetic coding is employed, each refinement symbol is coded using one of three contexts. The particular context employed is selected based on if the second MSB position is being refined and the significance of 8-connected neighbors. A.3.10.3 Cleanup Pass The third (and final) coding pass for each bit plane is the cleanup pass. This pass is used to convey significance and (as necessary) sign information for those samples that have not yet been found to be significant and are predicted to remain insignificant during the processing of the current bit plane. Appendix A . JPEG 2000: A n International Standard for Still Image Compression 131 i'i'"' /{•;' j Figure A. 10: Sample scan order within a code block. 1 for each vertical scan in code block do 2 if four samples in vertical scan and all previously insignificant and unvisited and none have significant 8-connected neighbor then 3 code number of leading insignificant samples via aggregation 4 skip over any samples indicated as insignificant by aggregation 5 endif 6 while more samples to process in vertical scan do 7 if sample previously insignificant and unvisited then 8 code significance of sample if not already implied by run /* 1 binary symbol */ 9 if sample significant then 10 code sign of sample /* 1 binary symbol */ 11 endif 12 endif 13 endwhile 14 endfor Algorithm A.3: Cleanup pass algorithm. Conceptually, the cleanup pass is not much different from the significance pass. The key difference is that the cleanup pass conveys information about samples that are predicted to remain insignificant, rather than those that are predicted to become significant. Algorithmically, however, there is one important difference between the cleanup and significance passes. In the case of the cleanup pass, samples are sometimes processed in groups, rather than individually as with the significance pass. Recall the scan pattern for samples in a code block, shown earlier in Figure A. 10. A code block is partitioned into stripes with a nominal height of four samples. Then, stripes are scanned from top to bottom, and the columns within a stripe are scanned from left to right. For convenience, we will refer to each column within a stripe as a vertical scan. That is, each vertical arrow in the diagram corresponds to a so called vertical scan. As will soon become evident, the cleanup pass is best explained as operating on vertical scans (and not simply individual samples). The cleanup pass simply processes each of the vertical scans in order, with each vertical scan being processed as follows. If the vertical scan contains four samples (i.e., is a full scan), significance information is needed for all of these samples, and all of the samples are predicted to remain insignificant, a special mode, called aggregation mode, is entered. In this mode, the number of leading insignificant samples in the vertical scan is coded. Then, the samples whose significance information is conveyed by aggregation are skipped, and processing continues with the remaining samples of the vertical scan exactly as is done in the significance pass. In pseudocode form, this process is described by Algorithm A.3. When aggregation mode is entered, the four samples of the vertical scan are examined. If all four samples are insignificant, an all-insignificant aggregation symbol is coded, and processing of the vertical scan is complete. Oth-erwise, a some-significant aggregation symbol is coded, and two binary symbols are then used to code the number of leading insignificant samples in the vertical scan. The symbols generated during the cleanup pass are always arithmetically coded. In the aggregation mode, the aggregation symbol is coded using a single context, and the two symbol run length is coded using a single context with a fixed uniform probability distribution. When aggregation mode is not employed, significance and sign coding Appendix A . JPEG 2000: An International Standard for Still Image Compression 132 function just as in the case of the significance pass. A.3.11 Tier-2 Coding In the encoder, tier-1 encoding is followed by tier-2 encoding. The input to the tier-2 encoding process is the set of bit-plane coding passes generated during tier-1 encoding. In tier-2 encoding, the coding pass information is packaged into data units called packets, in a process referred to as packetization. The resulting packets are then output to the final code stream. The packetization process imposes a particular organization on coding pass data in the output code stream. This organization facilitates many of the desired codec features including rate scalability and progressive recovery by fidelity or resolution. A packet is nothing more than a collection of coding pass data. Each packet is comprised of two parts: a header and body. The header indicates which coding passes are included in the packet, while the body contains the actual coding pass data itself. In the code stream, the header and body may appear together or separately, depending on the coding options in effect. Rate scalability is achieved through (quality) layers. The coded data for each tile is organized into L layers, num-bered from 0 to L— 1, where L > 1. Each coding pass is either assigned to one of the L layers or discarded. The coding passes containing the most important data are included in the lower layers, while the coding passes associated with finer details are included in higher layers. During decoding, the reconstructed image quality improves incrementally with each successive layer processed. In the case of lossy compression, some coding passes may be discarded (i.e., not included in any layer) in which case rate control must decide which passes to include in the final code stream. In the lossless case, all coding passes must be included. If multiple layers are employed (i.e., L > 1), rate control must decide in which layer each coding pass is to be included. Since some coding passes may be discarded, tier-2 coding is the second primary source of information loss in the coding path. Recall, from Section A.3.9, that each coding pass is associated with a particular component, resolution level, subband, and code block. In tier-2 coding, one packet is generated for each component, resolution level, layer, and precinct 4-tuple. A packet need not contain any coding pass data at all. That is, a packet can be empty. Empty packets are sometimes necessary since a packet must be generated for every component-resolution-layer-precinct combination even if the resulting packet conveys no new information. As mentioned briefly in Section A.3.7, a precinct is essentially a grouping of code blocks within a subband. The precinct partitioning for a particular subband is derived from a partitioning of its parent LL band (i.e., the LL band at the next higher resolution level). Each resolution level has a nominal precinct size. The nominal width and height of a precinct must be a power of two, subject to certain constraints (e.g., the maximum width and height are both 2 1 5). The LL band associated with each resolution level is divided into precincts. This is accomplished by overlaying the LL band with a regular grid having horizontal and vertical spacings of 2 m and 2PPy, respectively, as shown in Figure A . l 1, where the grid is aligned with the origin of the L L band's coordinate system. The precincts bordering on the edge of the subband may have dimensions smaller than the nominal size. Each of the resulting precinct regions is then mapped into its child subbands (if any) at the next lower resolution level. This is accomplished by using the coordinate transformation («, v) = ([*/2], [y/2]) where (x,y) and (u,v) are the coordinates of a point in the LL band and child subband, respectively. Due to the manner in which the precinct partitioning is performed, precinct boundaries always align with code block boundaries. Some precincts may also be empty. Suppose the nominal code block size is 2 x c b ' x 2 y c b ' . This results nominally in 2 P P x ' ~ x c b ' x 2ppy'~>'cb' groups of code blocks in a precinct, where and r is the resolution level. Since coding pass data from different precincts are coded in separate packets, using smaller precincts reduces the amount of data contained in each packet. If less data is contained in a packet, a bit error is likely to result in less information loss (since, to some extent, bit errors in one packet do not affect the decoding of other packets). Thus, using a smaller precinct size leads to improved error resilience, while coding efficiency is degraded due to the increased overhead of having a larger number of packets. for r = 0 for r > 0, for r = 0 for r > 0. (A. 12) (A. 13) Appendix A . JPEG 2000: An International Standard for Still Image Compression 133 (0,0) (m2 , n 2 I :: :::!: : : : ( t r x 0 , tr> K o 0) K , * 3 (trx ,-i K g If —™— J PPy PPy PPy Figure A . 11: Partitioning of a resolution into precincts. More than one ordering of packets in the code stream is supported. Such orderings are called progressions. There are five builtin progressions denned: 1) layer-resolution-component-position ordering, 2) resolution-layer-component-position ordering, 3) resolution-position-component-layer ordering, 4) position-component-resolution-layer ordering, and 5) component-position-resolution-layer ordering. The sort order for the packets is given by the name of the ordering, where position refers to precinct number, and the sorting keys are listed from most significant to least significant. For example, in the case of the first ordering given above, packets are ordered first by layer, second by resolution, third by component, and last by precinct. This corresponds to a progressive recovery by fidelity scenario. The second ordering above is associated with progressive recovery by resolution. The three remaining orderings are somewhat more esoteric. It is also possible to specify additional user-defined progressions at the expense of increased coding overhead. In the simplest scenario, all of the packets from a particular tile appear together in the code stream. Provisions exist, however, for interleaving packets from different tiles, allowing further flexibility on the ordering of data. If, for example, progressive recovery of a tiled image was desired, one would probably include all of the packets associated with the first layer of the various tiles, followed by those packets associated with the second layer, and so on. In the decoder, the tier-2 decoding process extracts the various coding passes from the code stream (i.e., depack-etization) and associates each coding pass with its corresponding code block. In the lossy case, not all of the coding passes are guaranteed to be present since some may have been discarded by the encoder. In the lossless case, all of the coding passes must be present in the code stream. In the sections that follow, we describe the packet coding process in more detail. For ease of understanding, we choose to explain this process from the point of view of the encoder. The decoder algorithms, however, can be trivially deduced from those of the encoder. A.3.11.1 Packet Header Coding The packet header corresponding to a particular component, resolution level, layer, and precinct, is encoded as follows. First, a single binary symbol is encoded to indicate if any coding pass data is included in the packet (i.e., if the packet is non-empty). If the packet is empty, no further processing is required and the algorithm terminates. Otherwise, we proceed to examine each subband in the resolution level in a fixed order. For each subband, we visit the code blocks belonging to the precinct of interest in raster scan order as shown in Figure A . 12. To process a single code block, we begin by determining if any new coding pass data is to be included. If no coding pass data has yet been included for this code block, the inclusion information is conveyed via a quadtree-based coding procedure. Otherwise, a binary symbol is emitted indicating the presence or absence of new coding pass data for the code block. If no new coding Appendix A . JPEG 2000: A n International Standard for Still Image Compression 134 Figure A. 12: Code block scan order with a precinct. l: if packet not empty then 2: code non-empty packet indicator /* 1 binary symbol */ 3: for each subband in resolution level do 4: for each code block in subband precinct do 5: code inclusion information /* 1 binary symbol or tag tree */ 6: if no new coding passes included then 7: skip to next code block 8: endif 9: if first inclusion of code block then 10: code number of leading insignificant bit planes /* tag tree */ 11: endif 12: code number of new coding passes 13: code length increment indicator 14: code length of coding pass data 15: endfor 16: endfor 17: else 18: code empty packet indicator /* 1 binary symbol */ 19: endif 20: pad to byte boundary Algorithm A.4: Packet header coding algorithm. passes are included, we proceed to the processing of the next code block in the precinct. Assuming that new coding pass data are to be included, we continue with our processing of the current code block. If this is the first time coding pass data have been included for the code block, we encode the number of leading insignificant bit planes for the code block using a quadtree-based coding algorithm. Then, the number of new coding passes, and the length of the data associated with these passes is encoded. A bit stuffing algorithm is applied to all packet header data to ensure that certain bit patterns never occur in the output, allowing such patterns to be used for error resilience purposes. The entire packet header coding process is summarized by Algorithm A.4. A.3.11.2 Packet Body Coding The algorithm used to encode the packet body is relatively simple. The code blocks are examined in the same order as in the case of the packet header. If any new passes were specified in the corresponding packet header, the data for these coding passes are concatenated to the packet body. This process is summarized by Algorithm A.5. A.3.12 Rate Control In the encoder, rate control can be achieved through two distinct mechanisms: 1) the choice of quantizer step sizes, and 2) the selection of the subset of coding passes to include in the code stream. When the integer coding mode is used (i.e., when only ITI transforms are employed) only the first mechanism may be used, since the quantizer step sizes must be fixed at one. When the real coding mode is used, then either or both of these rate control mechanisms may be employed. When the first mechanism is employed, quantizer step sizes are adjusted in order to control rate. As the step sizes Appendix A . JPEG 2000: A n International Standard for Still Image Compression 135 4: 2: 7: 3: 5: 6: for each subband in resolution level do for each code block in subband precinct do if (new coding passes included for code block) then output coding pass data endif endfor endfor Algorithm A.5: Packet body coding algorithm. are increased, the rate decreases, at the cost of greater distortion. Although this rate control mechanism is concep-tually simple, it does have one potential drawback. Every time the quantizer step sizes are changed, the quantizer indices change, and tier-1 encoding must be performed again. Since tier-1 coding requires a considerable amount of computation, this approach to rate control may not be practical in computationally-constrained encoders. When the second mechanism is used, the encoder can elect to discard coding passes in order to control the rate. The encoder knows the contribution that each coding pass makes to rate, and can also calculate the distortion reduction associated with each coding pass. Using this information, the encoder can then include the coding passes in order of decreasing distortion reduction per unit rate until the bit budget has been exhausted. This approach is very flexible in that different distortion metrics can be easily accommodated (e.g., mean squared error, visually weighted mean squared error, etc.). For a more detailed treatment of rate control, the reader is referred to [72] and [132]. A.3.13 Region of Interest Coding The codec allows different regions of an image to be coded with differing fidelity. This feature is known as region-of-interest (ROI) coding. In order to support ROI coding, a very simple yet flexible technique is employed as described below. When an image is synthesized from its transform coefficients, each coefficient contributes only to a specific region in the reconstruction. Thus, one way to code a ROI with greater fidelity than the rest of the image would be to identify the coefficients contributing to the ROI, and then to encode some or all of these coefficients with greater precision than the others. This is, in fact, the basic premise behind the ROI coding technique employed in the JPEG-2000 codec. When an image is to be coded with a ROI, some of the transform coefficients are identified as being more important than the others. The coefficients of greater importance are referred to as ROI coefficients, while the remaining coef-ficients are known as background coefficients. Noting that there is a one-to-one correspondence between transform coefficients and quantizer indices, we further define the quantizer indices for the ROI and background coefficients as the ROI and background quantizer indices, respectively. With this terminology introduced, we are now in a position to describe how ROI coding fits into the rest of the coding framework. The ROI coding functionality affects the tier-1 coding process. In the encoder, before the quantizer indices for the various subbands are bit-plane coded, the ROI quantizer indices are scaled upwards by a power of two (i.e., by a left bit shift). This scaling is performed in such a way as to ensure that all bits of the ROI quantizer indices lie in more significant bit planes than the potentially nonzero bits of the background quantizer indices. As a consequence, all information about ROI quantizer indices will be signalled before information about background ROI indices. In this way, the ROI can be reconstructed at a higher fidelity than the background. Before the quantizer indices are bit-plane coded, the encoder examines the background quantizer indices for all of the subbands looking for the index with the largest magnitude. Suppose that this index has its most significant bit in bit position N — 1. All of the ROI indices are then shifted N bits to the left, and bit-plane coding proceeds as in the non-ROI case. The ROI shift value N is included in the code stream. During decoding, any quantizer index with nonzero bits lying in bit plane N or above can be deduced to belong to the ROI set. After the reconstructed quantizer indices are obtained from the bit-plane decoding process, all indices in the ROI set are then scaled down by a right shift of N bits. This undoes the effect of the scaling on the encoder side. The ROI set can be chosen to correspond to transform coefficients affecting a particular region in an image or subset of those affecting the region. This ROI coding technique has a number of desirable properties. First, the ROI Appendix A . JPEG 2000: An International Standard for Still Image Compression 136 Type Length Parameters (if required) (if required) 16 bits 16 bits variable length Figure A. 13: Marker segment structure. • Main Header SOC Marker Segment | SIZ Marker Segment Other Marker Segments : (e.g., COD, COC, QCD, QCC, RGN, etc.) : Tile-Part Header | SOT Marker Segment ~ Other Marker Segments • (e.g., COD, COC, QCD, QCC, RGN, etc.) SOD Marker Segment • file-Part Body | Packet Data : Main Trailer ; EOC Marker Segment Figure A. 14: Code stream structure. can have any arbitrary shape and be disjoint. Second, there is no need to explicitly signal the ROI set, since it can be deduced by the decoder from the ROI shift value and the magnitude of the quantizer indices. For more information on ROI coding, the reader is referred to [46,25]. A.3.14 Code Stream In order to specify the coded representation of an image, two different levels of syntax are employed by the codec. The lowest level syntax is associated with what is referred to as the code stream. The code stream is essentially a sequence of tagged records and their accompanying data. The basic building block of the code stream is the marker segment. As shown in Figure A.13, a marker segment is comprised of three fields: the type, length, and parameters fields. The type (or marker) field identifies the particular kind of marker segment. The length field specifies the number of bytes in the marker segment. The parameters field provides additional information specific to the marker type. Not all types of marker segments have length and parameters fields. The presence (or absence) of these fields is determined by the marker segment type. Each type of marker segment signals its own particular class of information. A code stream is simply a sequence of marker segments and auxiliary data (i.e., packet data) organized as shown in Figure A . M . The code stream consists of a main header, followed by tile-part header and body pairs, followed by a main trailer. A list of some of the more important marker segments is given in Table A.2. Parameters specified in marker segments in the main header serve as defaults for the entire code stream. These default settings, however, may be overridden for a particular tile by specifying new values in a marker segment in the tile's header. All marker segments, packet headers, and packet bodies are a multiple of 8 bits in length. As a consequence, all markers are byte-aligned, and the code stream itself is always an integral number of bytes. Appendix A . JPEG 2000: An International Standard for Still Image Compression 137 Table A.2: Types of marker segments Type Description Start of codestream (SOC) Signals the start of a code stream. Always the first marker segment in the code stream (i.e., the first marker segment in the main header). End of codestream (EOC) Signals the end of the code stream. Always the last marker segment in the code stream. Start of tile-part (SOT) Indicates the start of a tile-part header. Always the first marker segment in a tile-part header. Start of data (SOD) Signals the end of the tile-part header. Always the last marker segment in the tile-part header. The tile body follows immediately after this marker segment. Image and tile size (SIZ) Conveys basic image characteristics (e.g., image size, number of compo-nents, precision of sample values), and tiling parameters. Always the second marker segment in the code stream. Coding style default (COD) Specifies coding parameters (e.g., multicomponent transform, wavelet/subband transform, tier-l/tier-2 coding parameters, etc.). Coding style component (COC) Specifies a subset of coding parameters for a single component. Quantization default (QCD) Specifies quantization parameters (i.e., quantizer type, quantizer parameters). Quantization component (QCC) Specifies quantization parameters for a single component. Region of interest (RGN) Specifies region-of-interest coding parameters. LBox TBox XLBox DBox (if required) 32 bits 32 bits 64 bits variable Figure A.15: Box structure. A.3.15 File Format A code stream provides only the most basic information required to decode an image (i.e., sufficient information to deduce the sample values of the decoded image). While in some simple applications this information is sufficient, in other applications additional data is required. To display a decoded image, for example, it is often necessary to know additional characteristics of an image, such as the color space of the image data and opacity attributes. Also, in some situations, it is beneficial to know other information about an image (e.g., ownership, origin, etc.) In order to allow the above types of data to be specified, an additional level of syntax is employed by the codec. This level of syntax is referred to as the file format. The file format is used to convey both coded image data and auxiliary information about the image. Although this file format is optional, it undoubtedly will be used extensively by many applications, particularly computer-based software applications. The basic building block of the file format is referred to as a box. As shown in Figure A. 15, a box is nominally comprised of four fields: the LBox, TBox, XLBox, and DBox fields. The LBox field specifies the length of the box in bytes. The TBox field indicates the type of box (i.e., the nature of the information contained in the box). The XLBox field is an extended length indicator which provides a mechanism for specifying the length of a box whose size is too large to be encoded in the length field alone. If the LBox field is 1, then the XLBox field is present and contains the true length of the box. Otherwise, the XLBox field is not present. The DBox field contains data specific to the particular box type. Some types of boxes may contain other boxes as data. As a matter of terminology, a box that contains other boxes in its DBox field is referred to as a superbox. Several of the more important types of boxes are listed in Table A.3. A file is a sequence of boxes. Since certain types of boxes are defined to contain others, there is a natural hierarchi-cal structure to a file. The general structure of a file is shown in Figure A. 16. The JPEG-2000 signature box is always first, providing an indication that the byte stream is, in fact, correctly formatted. The file type box is always second, indicating the version of the file format to which the byte stream conforms. Although some constraints exist on the ordering of the remaining boxes, some flexibility is also permitted. The header box simply contains a number of other Appendix A . JPEG 2000: A n International Standard for Still Image Compression 138 JPEG-2000 Signature Box File Type Box JP2 Header Box Image Header Box Color Specification Box Contiguous Code Stream Box Figure A. 16: File format structure. Table A.3: Box types Type Description JPEG-2000 Signature Identifies the file as being in the JP2 format. Always the first box in a JP2file. File Type Specifies the version of the format to which the file conforms. Always the second box in a JP2 file. JP2 Header Specifies information about the image aside from the coded image data itself. (A superbox.) Image Header Specifies the size and other basic characteristics of the image. Color Specification Specifies the colorspace to which the image sample data belongs. Contiguous Code Stream Contains a code stream. boxes. The image header box specifies several basic characteristics of the image (including image size, number of components, etc.). The bits per component box indicates the precision and signedness of the component samples. The color specification box identifies the color space of image data (for display purposes) and indicates which components map to which type of spectral information (i.e., the correspondence between components and color/opacity planes). Every file must contain at least one contiguous code stream box. (Multiple contiguous code stream boxes are permitted in order to facilitate the specification of image sequences to support trivial animation effects.) Each contiguous code stream box contains a code stream as data. In this way, coded image data is embedded into a file. In addition to the types of boxes discussed so far, there are also box types for specifying the capture and display resolution for an image, palette information, intellectual property information, and vendor/application-specific data. Although some of the information stored at the file format level is redundant (i.e., it is also specified at the code stream level), this redundancy allows trivial manipulation of files without any knowledge of the code stream syntax. The file name extension "jp2" is to be used to identify files containing data in the JP2 file format. For more information on the file format, the reader is referred to [66]. A.3.16 Extensions Although the baseline codec is quite versatile, there may be some applications that could benefit from additional fea-tures not present in the baseline system. To this end, Part 2 of the standard [73] is to define numerous extensions to the baseline codec. Some of the extensions likely to be defined include the following: 1) independent regions, a generalization of tiling whereby an image may be partitioned into arbitrarily-shaped and possibly overlapped regions; 2) additional intercomponent transforms (e.g., multidimensional wavelet/subband transforms); 3) additional intracom-ponent transforms (e.g., subband transforms based on arbitrary filters and decomposition trees, different filters in the horizontal and vertical directions); 4) overlapped transforms; 5) additional quantization methods such as trellis coded quantization [84,97]; 6) enhanced ROI support (e.g., a mechanism for explicitly signalling the shape of the ROI and an Appendix A . JPEG 2000: A n International Standard for Still Image Compression 139 Table A.4: Test images Image Characteristics Description bike grayscale, 2048x2560, 8 bpp/component grayscale version of bike_rgb image bike_rgb RGB color, 2048x2560, 8 bpp/component collection of objects (e.g., bike, fruit, newspaper, etc.) gold grayscale, 720x576, 8 bpp/component houses, road, and countryside target grayscale, 512x512, 8 bpp/component patterns and textures woman grayscale, 2048x2560, 8 bpp/component grayscale version of woman_rgb image woman_rgb RGB color, 2048x2560, 8 bpp/component lady (head and torso) arbitrary shift value); and 7) extensions to the file format including support for additional color spaces and compound documents. A . 4 Code c E v a l u a t i o n Having examined the details of the JPEG-2000 codec, we now evaluate the codec's performance. We begin by studying the computational complexity of the various operations in the encoder and decoder. Then, we compare JPEG-2000 to several other well-known compression methods. A.4.1 Evaluation Methodology In the sections that follow, we present experimental results obtained with a number of software packages. Here, we identify the particular software used, for the benefit of the reader. In the case of JPEG-2000, results were obtained with two different implementations: JasPer (version 0.016) [14,2,3], and the JPEG-2000 verification model (VM) software (version 7.1) [113]. For JPEG, the code from the Independent JPEG Group [61] (version 6b) was employed. In the case of JPEG LS, the software from the University of British Columbia (version 2.2) [ 1.03] was used. Finally, in the case of SPIHT, the software from the Rensselaer Polytechnic Institute was employed [124,115]. Al l of the above software are written in the C/C++ programming language. The experimental results presented in the manuscript were gathered on a Sun UltraSPARC machine (with a clock speed of 300 MHz and 512 MB of memory) running Solaris 5.6. The GNU C/C++ compiler (version 2.7.2) was used. For evaluation purposes, a suite of six test images were used, as listed in Table A.4. All of these images are taken from the JPEG-2000 test set [68]. A.4.2 Code Execution Profiling When implementing a complex algorithm in software, it is often beneficial to know which parts of the algorithm are likely to contribute the most to total execution time. With this as motivation, we have performed some basic profiling of the codec software, the results of which are presented below. For evaluation purposes, three of the test images listed in Table A.4 were employed: bike, bike_rgb, and target. This test suite was chosen to include grayscale, color, natural, and synthetic imagery. Two typical coding scenarios were considered: one lossless and one lossy. In each case, we measured the execution time associated with each of the basic encoder/decoder operations (i.e., multicomponent transform, wavelet transform, quantization, tier-1 coding, tier-2 coding, and rate control). Appendix A . JPEG 2000: An International Standard for Still Image Compression 140 A.4.2.1 Lossless Coding In the first coding scenario, the three test images were encoded in a lossless manner, and then the resulting code streams were decoded. In each case, the encoder and decoder execution was profiled, leading to the results shown in Table A.5. First, let us consider the case of the encoder. Examining Table A.5(a), we see that the encoder spends most of its time on tier-1 coding and the wavelet transform. In fact, these two operations together account for the vast majority of the encoder's execution time (typically 85-95%). The amount of time consumed by the multicomponent transform and tier-2 coding operations is much less significant. Examining tier-1 coding in more detail, we observe that for natural imagery (e.g., bike, bike_rgb), significance coding usually requires the most time, followed by the cleanup coding, and then refinement coding. For synthetic imagery, however, this trend is not necessarily observed, as evidenced by the results for the target image. Thus, the execution time breakdown for tier-1 coding is somewhat sensitive to image content. Now, let us consider the case of the decoder. The execution profiling results for the decoder are given in Ta-ble A.5(b). Again, we can see that the decoder usually spends most of its time on tier-1 coding (typically about 50%), followed next by the wavelet transform. These two operations combined account for the vast majority of the decoder's execution time (typically 80-95%). The multicomponent transform and tier-2 coding operations consume relatively little time. Examining the tier-1 coding process in more detail, we observe similar trends as in the encoder. That is, for natural imagery, significance coding requires the most time, followed by cleanup coding, and then refinement coding, while the breakdown is less predictable for synthetic imagery. A.4.2.2 Lossy Coding In the second coding scenario, the three test images were encoded in a lossy manner, and then the resulting code streams were decoded. Again, in each case, the encoder and decoder execution was profiled. This led to the results given in Table A.6. First, let us study the results for the encoder, given in Table A.6(a). Clearly, the tier-1 coding and wavelet transform operations require the most time. Together, these two operations account for the vast majority of the execution time (typically 75-85%). The multicomponent transform is the next biggest consumer of time (typically, requiring 10-15%). The quantization, tier-2 coding, and rate control operations require relatively little time. Looking at the tier-1 coding process in more detail, we observe that the most time is usually required by cleanup coding, followed by significance coding, and then refinement coding. Now, let us consider the case of the decoder. From the results in Table A.6(b), we can see that the wavelet transform clearly requires the most time by far (typically 45-70%). Tier-1 coding is the next largest time consumer, but has much more modest requirements (typically 15-25%). The multicomponent transform requires almost as much time (typically 15-20%). The quantization, and tier-2 coding operations require relatively little time. Looking at the tier-1 coding process in more detail, we observe that, for natural imagery, the least time is required by refinement coding. Generally speaking, the execution time breakdown for tier-1 coding depends, to a large extent, on the particular image and bit rate employed. In the case of both JPEG-2000 implementations considered in this study (i.e., JasPer and the V M software), rate control is achieved through the selection of the coding passes to include in the final code stream. The quantizer step sizes are fixed. To ensure optimal rate-distortion performance, the encoder generates all of the coding passes during tier-1 coding (even though some coding passes may not be included in the final code stream). This leads to the large amount of time consumed by tier-1 coding in the encoder. Since quantization takes place, however, the tier-1 coding overhead is not as great (in absolute terms) as in the lossless case. It is important to note that different multicomponent and wavelet transforms are employed in the lossy and lossless cases. In the case of lossless coding, the multicomponent and wavelet transforms employed are the RCT and 5/3, respectively. For lossy coding, the ICT and 9/7 transforms are employed, respectively. The ICT requires more compu-tation than the RCT, and similarly the 9/7 transform requires more computation than the 5/3 transform. This explains the larger amount of time spent on the multicomponent and wavelet transform operations relative to the lossless case. A.4.3 JPEG 2000 vs. Other Methods So far, we have examined the JPEG 2000 codec only in isolation. Of course, it is only natural to wonder how JPEG 2000 performs in comparison to other compression methods. With this as motivation, we compare the performance Appendix A . JPEG 2000: A n International Standard for Still Image Compression 141 Table A.5: Codec execution profile for lossless coding (1 tile, SNR progressive, 1 layer, RCT for color images, 5/3 wavelet transform, 5 decomposition levels), (a) Encoder, (b) Decoder. (a) % of Execution Time Operation bike bike_rgb target intercomponent transform — 0.5 — intracomponent transform 40.4 42.9 46.4 tier-1 coding 52.8 49.1 46.3 significance 22.0 21.5 12.9 refinement 13.1 10.5 9.5 cleanup 14.4 13.6 19.5 tier-2 coding 4.1 5.0 2.6 (b) % of Execution Time Operation bike bike..._rgb target intercomponent transform — 0.5 — intracomponent transform 40.8 41.2 52.2 tier-1 coding 51.3 45.2 39.2 significance 19.9 18.3 8.9 refinement 9.5 8.3 11.2 cleanup 11.3 8.7 6.6 tier-2 coding 3.1 8.1 1.6 Table A.6: Codec execution profile for lossy coding (32:1 compression for grayscale images and 64:1 compression for color images, 1 tile, SNR progressive, 1 layer, ICT for color images, 9/7 wavelet transform, 5 decomposition levels), (a) Encoder, (b) Decoder. (a) % of Execution Time Operation bike bike_rgb target intercomponent transform — 12.7 — intracomponent transform 31.7 29.9 35.9 quantization 3.1 2.8 3.4 tier-1 coding 60.7 49.7 54.1 significance 22.5 17.8 16.6 refinement 12.5 8.6 13.5 cleanup 19.9 16.8 19.6 tier-2 coding 1.3 1.1 0.7 (b) % of Execution Time Operation bike bike_rgb target intercomponent transform — 18.5 — intracomponent transform 63.3 48.6 58.9 quantization 6.3 4.8 5.4 tier-1 coding 19.6 17.4 20.2 significance 3.7 2.4 3.6 refinement 1.1 0.6 1.5 cleanup 3.3 2.2 6.0 tier-2 coding 0.5 0.2 0.3 Appendix A . JPEG 2000: A n International Standard for Still Image Compression 142 Table A.7: Comparison of JPEG 2000 and other methods for lossless coding NBR T Encoder Time (s) Decoder Time (s) JPEG JPEG JPEG JPEG JPEG JPEG Image 2000 SPIHT LS 2000 SPIHT LS 2000 SPIHT LS bike 0.556 0.560 0.544 16.19 30.27 4.45 14.70 30.84 3.70 bike_rgb 0.498 — 0.560 45.12 — 19.67 45.56 — 16.08 gold 0.575 0.577 0.560 1.26 2.28 0.37 1.16 2.33 0.31 target 0.266 0.331 0.273 0.63 1.12 0.16 0.55 1.15 0.13 woman 0.564 0.552 0.556 15.80 29.66 4.48 14.39 30.36 3.71 woman_.rgb 0.479 — 0.572 43.25 — 19.75 43.95 — 16.08 Mean* 0.490 0.505 0.510 ^The normalized bit rate (NBR) is defined as the reciprocal of the compression ratio. *The mean taken over all of the grayscale images. of JPEG 2000 to several other well-known compression techniques, namely JPEG, JPEG LS, and SPIHT. JPEG, of course, is the welUknown international standard most frequently employed for lossy compression. JPEG LS is a relatively new international standard for lossless (and near-lossless) compression. And lastly, SPIHT is a de facto benchmark used within the academic community. For lossless coding, we compare JPEG 2000 to SPIHT and JPEG LS, while for lossy coding, we compare JPEG 2000 to SPIHT and JPEG. Each of the test images listed in Table A.4 was compressed in a lossless manner and then decompressed using the JPEG-2000, SPIHT, and JPEG LS codecs. In each case, the bit rate, encoder execution time, and decoder execution time was measured, yielding the results in Table A.7. Three of the test images from Table A.4 (i.e. bike, bike_rgb, and target) were compressed in a lossy manner at several bit rates, and then decompressed using the JPEG-2000, SPIHT, and JPEG codecs. In each case, the distortion, encoder execution time, and decoder execution time was measured, yielding the results in Table A.8. In the sections that follow, we examine the above results in detail. A.4.3.1 JPEG 2000 vs. JPEG LS We begin by comparing the lossless coding results for JPEG 2000 and JPEG LS in Table A.7. For RGB images, JPEG 2000 achieves significantly higher compression (typically 10-15% better). For other types of imagery, JPEG LS usually fares better, although the performance gain is frequently marginal (typically 1-2%). Examining the timing results, we see that the JPEG-2000 encoder is slower than the JPEG-LS encoder by a factor of about 2.25-4.0, with the factor tending to be smaller for RGB images. The JPEG-2000 decoder is slower than the JPEG-LS decoder by factor of approximately 2.5-4.5, again, with the factor normally being smaller for RGB images. JPEG 2000 utilizes a multicomponent transform to reduce intercomponent redundancies, while JPEG-LS does not. This is, no doubt, a contributing factor to the significantly better compression efficiency achieved by JPEG 2000 on RGB imagery. In the case of both encoding and decoding, the relative difference in execution time between JPEG 2000 and JPEG LS is smaller for RGB imagery. This behavior can be attributed to the higher compression efficiency obtained by JPEG 2000 which results in less I/O overhead and a relative decrease in execution time. Clearly, the main advantage that JPEG LS has over JPEG 2000 is execution speed. This is to be expected, however. With JPEG LS, no transform is used, and a relatively simple entropy coding technique is employed (i.e., run-length and Golomb coding). On the other hand, JPEG 2000 employs multicomponent/wavelet transforms and complex bit-plane and entropy coding techniques (i.e., context-based adaptive arithmetic coding). In terms of compression efficiency, there is no clear winner. For some classes of imagery, JPEG 2000 yields superior results, while for others JPEG LS fares better. Computational complexity aside, JPEG 2000 has a number of important advantages over JPEG LS: 1) JPEG 2000 provides an efficient framework for both lossy and lossless compression, whereas JPEG LS only sup-ports lossless (or near-lossless) compression; 2) JPEG 2000 is both rate and resolution scalable, while JPEG LS is neither. 3) For RGB imagery, JPEG 2000 has significantly better compression efficiency. Appendix A . JPEG 2000: A n International Standard for Still Image Compression 143 Table A.8: Comparison of JPEG 2000 and other methods for lossy coding. Lossy coding results for the (a) bike, (b) bike_rgb, and (c) target images. (a) PSNR (dB) Encoder Time (s) Decoder Time (s) • JPEG JPEG JPEG NBRt 2000 SPIHT JPEG 2000 SPIHT JPEG 2000 SPIHT JPEG 1 64 23.81 23.44 20.61 12.51 8.44 1.79 5.85 8.69 0.56 26.38 25.89 24.67 12.52 8.90 1.83 6.04 9.17 0.63 T ¥ Ifi 29.65 29.12 27.21 12.57 9.92 1.88 6.20 10.07 0.75 33.55 33.00 30.59 12.64 11.87 2.00 6.60 11.89 0.97 (b) PSNR (dB) Encoder Time (s) Decoder Time (s) JPEG JPEG JPEG NBR f 2000 SPIHT JPEG 2000 SPIHT JPEG 2000 SPIHT JPEG 1 2^ 6 T|5 (3 24.87,24.39,23.91 24.10,23.75,23.70 — 39.74 34.18 — 23.26 26.31 — 27.69,27.08,26.39 26.83,26.44,26.34 24.75,23.22,24.37 39.94 34.93 3.37 23.47 26.86 1.85 30.93,30.35,29.14 30.09,29.61,29.23 28.04,26.54,27.94 40.00 36.49 3.47 23.84 28.26 2.02 T 17 34.33,34.20,32.05 33.69,33.25,32.35 31.59,29.52,31.37 40.06 39.50 3.66 24.73 31.03 2.34 (c) PSNR (dB) Encoder Time (s) Decoder Time (s) JPEG JPEG JPEG NBRt 2000 SPIHT JPEG 2000 SPIHT JPEG 2000 SPIHT JPEG 1 20.45 16.62 — 0.56 0.34 — 0.37 0.35 r ? ¥ -re 23.78 20.40 18.16 0.55 0.37 0.11 0.34 0.37 0 04 27.99 24.27 20.50 0.56 0.42 0.11 0.34 0.41 0.06 34.90 29.90 26.15 0.56 0.51 0.11 0.33 0.49 0.06 tThe normalized bit rate (NBR) is defined as the reciprocal of the compression ratio. A.4.3.2 JPEG 2000 vs. JPEG Next, we compare the lossy coding results for JPEG 2000 and JPEG in Table A.8. As these results indicate, at low to medium bit rates, JPEG 2000 clearly outperforms JPEG in terms of PSNR. Subjectively, the distortion is also much less in the JPEG 2000 case, as evidenced by the example in Figure A.17. The improvement in coding efficiency offered by JPEG 2000 does not come without a cost, however. Over the range of compression considered in our study, the JPEG-2000 encoder is slower than the JPEG encoder by a factor of approximately 5.0-12.0, while the JPEG-2000 decoder is slower than the JPEG decoder by a factor of about 5.0-12.75. The particular factors depend on the bit rate considered and the type of imagery involved. As the bit rate decreases, the speed of JPEG-2000 relative to JPEG decreases. It should not be surprising that the JPEG-2000 codec is slower than the JPEG codec. JPEG-2000 requires more computation due to its use of bit-plane coding and more complex entropy coding (i.e., context-based adaptive arith-metic coding). As demonstrated above, JPEG 2000 offers vastly superior image quality at high compression, compared to JPEG. Although an important advantage, this is not the only one, however. JPEG 2000 also offers several other key advantages over JPEG. That is, JPEG 2000: 1) provides efficient lossless compression, unlike the lossless mode of JPEG; 2) offers better rate scalability, and 3) supports additional features/functionalities, such as region of interest coding, and a more flexible file format. A.4.3.3 JPEG 2000 vs. SPIHT Finally, we compare JPEG-2000 to SPIHT for both lossless and lossy compression. First, let us compare the lossless coding results for JPEG 2000 and SPIHT given in Table A.7. Since the SPIHT codec software does not support the lossless coding of multicomponent images, we only consider the single-component imagery here. In terms of coding efficiency, JPEG 2000 and SPIHT are roughly comparable, with JPEG 2000 perhaps having a slight edge. The JPEG-2000 encoder is faster than the SPIHT encoder by a factor of approximately 1.75-2.0, while the JPEG-2000 decoder is faster than the SPIHT decoder by a factor of about 2.0-2.25. Let us now compare the lossy coding results for JPEG 2000 and SPIHT given in Table A.8. From these numbers, Appendix A. JPEG 2000: An International Standard for Still Image Compression 144 (a) 0» Figure A. 17: Lossy compression example comparing JPEG 2000 and JPEG. The target image after 43:1 compression with (a) JPEG 2000 (PSNR of 26.10 dB) and (b) JPEG (PSNR of 18.91 dB). we can see that JPEG 2000 offers better PSNR performance than SPIHT. In terms of execution speed, the JPEG-2000 encoder is slower than the SPIHT encoder by a factor of about 1.0-1.75, while the JPEG-2000 decoder is faster than the SPIHT decoder by a factor of approximately 1.0-2.0. As the bit rate decreases, the speed of the JPEG-2000 encoder and decoder decrease relative to SPIHT. For lossless coding, the JPEG-2000 codec is faster than the SPIHT codec. This speed difference can be attributed to at least two factors. First, the JPEG-2000 codec utilizes a very fast multiplication-free arithmetic coder, while the SPIHT codec employs an arithmetic coder similar to the classic example given in [147]. Also, JPEG-2000 exploits symbol aggregation (in cleanup passes) to reduce the number of symbols that need to be arithmetically coded, while SPIHT does not utilize this type of strategy. For lossy coding, the JPEG-2000 decoder is faster than the SPIHT decoder. Again, this can be attributed to the faster entropy coding employed by JPEG 2000. The JPEG-2000 encoder tends to be slower than the SPIHT encoder, in spite of the fact that JPEG 2000 employs an inherently faster entropy coder. This is due to fact that JPEG 2000 incurs the expense of losslessly coding the quantized transform coefficients regardless of the target bit rate. A.5 Summary In this appendix, we commenced with a high-level introduction to the JPEG-2000 standard, and proceeded to study the JPEG-2000 codec in detail. We then studied the codec's computational complexity. As our code execution profiling results demonstrate, efficient wavelet transform and tier-1 coding engines are essential for a fast JPEG-2000 codec im-plementation. Lastly, we compared JPEG 2000 to various other well-known methods for compression (i.e., JPEG LS, JPEG, and SPIHT). Although requiring more CPU time than some methods, the JPEG-2000 offers superior coding ef-ficiency, and provides many new features/functionalities not available in existing standards. With its excellent coding performance and many attractive features, JPEG-2000 will no doubt become a widely used standard in the years to come. Appendix A . JPEG 2000: A n International Standard for Still Image Compression 145 A.6 JasPer The JasPer software is available from the JasPer Project web site (i.e., http: //ece.ubc.ca/~mdadams/jasper). This software has also been published by the ISO in the JPEG-2000 Part-5 standard as a reference implementation of the JPEG-2000 Part-1 codec. For more information about the JasPer software, the reader is referred to [2,14]. Acknowledgment The author would like to thank Dr. Hong Man for his assistance in collecting some of the experimental results reported in this appendix. As our own species is in the process of proving, one cannot have superior science and inferior morals. The combination is unstable and self-destroying. —Arthur C. Clarke 146 Index Symbols z transform 8 A aliasing 10 analysis filter bank 13 analysis filters 13 Aristotle 5 B biased ceiling function 7 biased floor function 7 biased truncation function 7 bit rate 20 block transform 8 C ceiling function 7 Clarke, Arthur Charles 109, 145 compression ratio 19 conditions for perfect reconstruction 16 coset 9 coset vector 9 D Dickens, Charles xiii distortion 20 downsampler 9 downsampling 9 E Einstein, Albert 117 F filter bank 13 fixed point arithmetic 2 floating point arithmetic 2 floor function 7 forward polyphase transform 16 G Gandhi, Mohandas Karamchand 60 generalized reversible ITI transform framework 33-44 Greene, Graham 20 H Haar wavelet transform 28 Huxley, Thomas Henry 107 I image coding 17 image compression 18 imaging 11 integer lattice 9 integer-bias invariance 8, 22 inverse polyphase transform 16 J JPEG 2000 118-145 L lattice 9 Lau, Evelyn 21, 86 lifting framework 30-32 M maximally decimated 13 mean-squared error 20 Mulder, Fox 61 multirate 9 N noble identities 11 nonexpansive transform 62 nonseparable sampling 9 normalized bit rate 20 O oddness 8,22 overlapping rounding transform framework — 32-33 P peak-signal-to-noise ratio 20 per-displace-step extension 64 perfect reconstruction 13 periodic extension 62 polyphase components 12 polyphase form of filter bank 14 polyphase representation — . — 11 R RAFZ function 7 Index 147 reversible 1 Rilke, Rainer Maria 6 s S transform 28-29 S+P transform framework 29-30 sampling matrix 9 Scully, Dana 108 separable sampling 9 shift-free perfect reconstruction 14 subband signals 13 sublattice 9 symmetric extension 62 synthesis filter bank 13 synthesis filters 13 T Tolkein, John Ronald Reuel 118 truncation function 7 U . uniformly decimated 13 uniformly maximally decimated 13 uniformly maximally-decimated filter bank 13 unimodular 6 upsampling 11 W Walpole, Horace 85 X X-Files, The 61, 108
- Library Home /
- Search Collections /
- Open Collections /
- Browse Collections /
- UBC Theses and Dissertations /
- Reversible interger-to-interger wavelet transforms...
Open Collections
UBC Theses and Dissertations
Featured Collection
UBC Theses and Dissertations
Reversible interger-to-interger wavelet transforms for image coding Adams, Michael David 2002
pdf
Page Metadata
Item Metadata
Title | Reversible interger-to-interger wavelet transforms for image coding |
Creator |
Adams, Michael David |
Date Issued | 2002 |
Description | Reversible integer-to-integer (ITI) wavelet transforms are studied in the context of image coding. Considered are matters such as transform frameworks, transform design techniques, the utility of transforms for image coding, and numerous practical issues related to transforms. The generalized reversible ITI transform (GRITIT) framework, a single unified framework for reversible ITI wavelet/block transforms, is proposed. This new framework is then used to study several previously proposed frameworks and their interrelationships. For example, the framework based on the overlapping rounding transform is shown to be a special case of the lifting framework with only trivial extensions. The applicability of the GRITIT framework for block transforms is also demonstrated. Throughout all of this work, particularly close attention is paid to rounding operators and their characteristics. Strategies for handling the transformation of arbitrary-length signals in a nonexpansive manner are considered (e.g., symmetric extension, per-displace-step extension). Two families of symmetry-preserving transforms (which are compatible with symmetric extension) are introduced and studied. We characterize transforms belonging to these families. Some new reversible ITI structures that are useful for constructing symmetry-preserving transforms are also proposed. A simple search-based design technique is explored as means for finding effective low-complexity transforms in the above-mentioned families. In the context of image coding, a number of reversible ITI wavelet transforms are compared on the basis of their lossy compression performance, lossless compression performance, and computational complexity. Of the transforms considered, several were found to perform particularly well, with the best choice for a given application depending on the relative importance of the preceding criteria. Reversible ITI versions of numerous transforms are also compared to their conventional (i.e., non-reversible real-to-real) counterparts for lossy compression. At low bit rates, reversible ITI and conventional versions of transforms were found to often yield results of comparable quality. Factors affecting the compression performance of reversible ITI wavelet transforms are also presented, supported by both experimental data and theoretical arguments. In addition to this work, the JPEG-2000 image compression standard is discussed. In particular, the JPEG-2000 Part-1 codec is described, analyzed, and evaluated. |
Extent | 15815637 bytes |
Genre |
Thesis/Dissertation |
Type |
Text |
FileFormat | application/pdf |
Language | eng |
Date Available | 2009-09-29 |
Provider | Vancouver : University of British Columbia Library |
Rights | For non-commercial purposes only, such as research, private study and education. Additional conditions apply, see Terms of Use https://open.library.ubc.ca/terms_of_use. |
IsShownAt | 10.14288/1.0065047 |
URI | http://hdl.handle.net/2429/13302 |
Degree |
Doctor of Philosophy - PhD |
Program |
Electrical and Computer Engineering |
Affiliation |
Applied Science, Faculty of Electrical and Computer Engineering, Department of |
Degree Grantor | University of British Columbia |
GraduationDate | 2002-11 |
Campus |
UBCV |
Scholarly Level | Graduate |
AggregatedSourceRepository | DSpace |
Download
- Media
- 831-ubc_2002-749835.pdf [ 15.08MB ]
- Metadata
- JSON: 831-1.0065047.json
- JSON-LD: 831-1.0065047-ld.json
- RDF/XML (Pretty): 831-1.0065047-rdf.xml
- RDF/JSON: 831-1.0065047-rdf.json
- Turtle: 831-1.0065047-turtle.txt
- N-Triples: 831-1.0065047-rdf-ntriples.txt
- Original Record: 831-1.0065047-source.json
- Full Text
- 831-1.0065047-fulltext.txt
- Citation
- 831-1.0065047.ris
Full Text
Cite
Citation Scheme:
Usage Statistics
Share
Embed
Customize your widget with the following options, then copy and paste the code below into the HTML
of your page to embed this item in your website.
<div id="ubcOpenCollectionsWidgetDisplay">
<script id="ubcOpenCollectionsWidget"
src="{[{embed.src}]}"
data-item="{[{embed.item}]}"
data-collection="{[{embed.collection}]}"
data-metadata="{[{embed.showMetadata}]}"
data-width="{[{embed.width}]}"
async >
</script>
</div>
Our image viewer uses the IIIF 2.0 standard.
To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.831.1-0065047/manifest