Sparse Signal Recovery in a Transform Domain by Evgeniy Lebed B.Sc., Simon Fraser University, 2006 A THESIS SUBMITTED IN PARTIAL FULFILMENT OF THE REQUIREMENTS FOR THE DEGREE OF Master of Science in The Faculty of Graduate Studies (Mathematics) The University Of British Columbia (Vancouver) August, 2008 © Evgeniy Lebed 2008 Abstract The ability to efficiently and sparsely represent seismic data is becoming an increasingly important problem in geophysics. Over the last thirty years many transforms such as wavelets, curvelets, contourlets, surfacelets, shearlets, and many other types of ‘x-lets’ have been developed. Such transform were leveraged to resolve this issue of sparse representations. In this work we compare the properties of four of these commonly used transforms, namely the shift-invariant wavelets, complex wavelets, curvelets and surfacelets. We also explore the performance of these transforms for the problem of recov ering seismic wavefields from incomplete measurements. II Table of Contents Abstract ii Table of Contents iii List of Tables v List of Figures vi Preface viii Acknowledgments ix Dedication x 1 Introduction 1 2 The four transforms 2.1 Motivation 2.2 Separable transforms 2.2.1 The shift-invariant wavelet transform 2.2.2 The dual-tree complex wavelet transform (CWT) 2.3 Non-separable transforms 2.3.1 The curvelet transform 2.3.2 The surfacelet transform 3 Sparse seismic signal recovery 3.1 Motivation 3.2 Model 3.3 Compressive sampling 3.4 £ minimization algorithm . . 3 3 4 4 7 8 8 10 13 13 14 15 17 H’ Table of Contents 4 Results 4.1 Recovering from a set of missing traces 4.2 Recovering from a set of missing angular frequencies 4.2.1 A different sampling scheme 4.2.2 The importance of random sampling 4.2.3 Direct Fourier spectra interpolation 4.3 Compressibility 4.3.1 Nonlinear approximation 4.3.2 The Pareto curve 4.4 Mutual coherence 4.4.1 Coherence approximation 4.4.2 Data-dependent coherence approximation 20 20 22 26 28 30 33 33 36 37 38 39 Conclusions and discussion 42 . 5 . . Appendices A Wavelet concepts A.1 Frames 44 44 Bibliography 45 iv List of Tables 4.1 4.2 4.3 4,4 4.5 5.1 Physical restriction performance summary Frequency restriction performance summary Coherence approximations Data-dependent coherence approximations using 10000 most significant transforms’ coefficients Data-dependent coherence approximations using 10000 most insignificant transforms’ coefficients Recovery performance summary 22 26 39 40 40 42 V List of Figures 2.1 2.2 2.3 Examples of seismic data sets. (a) real data (b) synthetic data 3 CWT filter bank (adapted from Selesnick et al. (2005)). Left: two levels of the analysis filter bank. Right: synthesis filter bank 7 Frequency domain partitioning of the curvelet transform (adapted from Candès et al. (2006)). One curvelet is strictly localized to one angular wedge (shaded) in the frequency domain. 9 (Adapted from Lu and Do (2007)) Left: first two level of the analysis filter bank of the surfacelet transform. Subsequent levels of decomposition are achieved by recursively inserting the point a into Right: synthesis filter bank of the surfacelet transform 11 . 2.4 3.1 3.2 4.1 A typical land seismic survey setup (adapted from www.iongeo.com). The survey consists of Vibroseis trucks and an array of geo phones spread around the survey area. There are physical restrictions on where sensors can be placed: e.g., one cannot put sensors on the roads or the lake 13 Example of the soft-thresholding function T, (x) with a thresh old level A = 3 18 Seismic data and missing traces. (a) Model synthetic seismic wavefield. (b) Data same seismic wavefield as in (a) but with 50% of the traces removed Recovered seismic wavefield via (a) curvelets; (b) surfacelets; (c) shift-invariant wavelets; (d) complex wavelets Fourier spectrum and missing frequencies. (a)The model fre quency spectrum of the synthetic wavefield in Figure 4.1(a). (b) The data 50% of the temporal frequencies are missing. The data from Figure 4.3(b) in the physical domain. Aliasing artifact are present throughout the entire image - - 4.2 4.3 21 - - 4.4 20 23 23 vi List of Figures 4.5 4.6 4.7 4.8 4.9 4.10 4.11 4.12 4.13 4.14 4.15 4.16 Recovered seismic wavefield via (a) curvelets; (b) surfacelets; (c) shift-invariant wavelets; (d) complex wavelets Fourier spectra of the recovered signals via (a) curvelets; (b) surfacelets; (c) shift-invariant wavelets; (d) complex wavelets. Frequency-weighted sampling scheme. (a) Averaged frequency distribution with 20% of the frequencies used to sample the signal (red lines); (b) recovered signal via shift-invariant wavelets with the frequency samples taken from (a) A 1D signal and it’s Fourier spectrum Randomly subsampled 1D signal and it’s Fourier spectrum Uniformly subsampled 1D signal and it’s Fourier spectrum. Uniform sub-sampling of angular frequencies. (a) Three-fold uniformly subsampled frequency spectra of the synthetic wavefield; (b) Attempted recovery result by shift invariant wavelets with a uniform subsampling scheme from (a); (c) interpolated spectrum Subsampled signal. (a) Subsampled Fourier spectrum of a seismic signal; (b) same signal as in (a), but in the physical domain Directly interpolating the Fourier spectra. (a) Interpolated spectra from Figure 4.12(a); (b) interpolated signal in the physical domain Nonlinear approximation. (a): Randomly sampled transform coefficients plotted in magnitude-descending order. (b) MSE of the reconstruction from an incomplete set of transform coefficients Pareto curves for the different transforms. The lined in the plot represent Pareto curves for compressibility of seismic data by: Fourier (yellow), curvelets (blue), surfacelets (red), complex wavelets (green) and shift-invariant wavelets (black) Single atoms of the four transforms in the physical and fre quency domains. (a): curvelet (top right), surfacelet (top right), shift-invariant wavelet (bottom left), complex wavelet (bottom right) in the physical domain. (b): Fourier transform of the atoms in the same order . . 24 25 27 28 29 30 31 32 32 35 37 41 vii Preface This thesis was prepared with Madagascar, a reproducible research soft-. ware package available at rsf.sf.net. Many results shown in this thesis are directly linked to the codes that generated them. This reproducibil ity allows direct access to the experiments preformed in this work allowing transparency and dissemination of knowledge for the Seismic Laboratory for Imaging and Modeling (SLIM), its sponsors, and the entire research com munity as a whole. The programs required to reproduce the reported results are Madagascar programs written in C/C++, Matlab, or Python. The numerical algorithms and applications are mainly written in Matlab with a few exceptions writ ten in Python as part of SLIMpy (slim.eos.ubc,ca/SLIMpy). The external Matlab toolboxes required to reproduce some of the results in this thesis are: Curvelab, Surfbox, RWT, and Complex wavelet toolbox. vii’ Acknowledgments First of all I would like to thank my supervisors, Ozgur Yilmaz and Felix Herrmann for the never-ending support and advice which they have provided me with throughout these last two years. Without their help I would not have been able to learn half the things which I have learned. Their patience with me is much appreciated. I would like to express my gratitude to Gilles Hennenfent, Ewout van den Berg, Cody Brown and the rest of the SLIM team for their help with resolving programming and other scientific issues. Just as well, I would like to thank Yue Lu for helping me unravel the thread of puzzles of the surfacelet transform. And lastly I would like to express my deepest thanks to all of my family and my lovely fiancé Kristina. There is no question that without your guys’ help I would not have been able to get to where I am today. ix To my grandfather Dr. Igor A. Karnovsky. x Chapter 1 Introduction There are many different transforms that can be used to obtain a sparse representations of signals (e.g., seismic images, natural images, and audio signals). Numerous ‘sparsifying’ transforms have been invented and exten sively studied in the seismic processing literature (see, e.g., Holschneider et al. (1989), Donoho and Johnstone (1994), Candès et al. (2006)). In this work we compare some of the important properties that these transforms possess. A sparsifying transform is an invertible linear operator, mapping a signal (e.g., an image) to a sparse vector, i.e., a vector with few large compo nents. Different types of images have sparse representations under different transforms. For example, piecewise constant images can be sparsely repre sented by spatial finite differences. Natural, real-life images are known to have a sparse representation in the discrete cosine transform (DCT) and wavelet transform domains (Dagher et al. (2003)). It has recently been ob served that the curvelet transform yields a sparse representation of seismic data (the curvelet dictionary consists of atoms that locally behave like waves, and seismic data, consisting of solutions to the wave equation, has a good sparse representation in curvelets; see, e.g., Candès and Demanet (2005)). Of course, in any practical setting, one does not expect to have a sparse representation with a non-adaptive transform. Therefore, strictly speaking, when we say “sparse” we mean that the magnitude-sorted transform coeffi cients decay rapidly. In this case we also say the signal is “compressible.” More precisely, the coefficient vectors live in a bounded ball in weak with 0< q <1. Surfacelets also try to achieve the same result, but with a much different construction. The complex-wavelet transform (Selesnick et al. (2005)) is of interest because it has a fixed and low redundancy factor and has the ability to give signals a multidirectional representation. Classical shift-invariant wavelets are of interest because computationally they provide a very fast multiscale representation of a signal although at a price of a high redundancy factor. Transform design is only one of the challenges towards obtaining sparse representations for signals. The second obstacle we need to tackle is find 1 Chapter 1. Introduction ing the sparse coefficients for a given signal. This is easy if the transform corresponds to an orthonormal basis expansion as there is a unique coeffi cient vector in this case. However, many transforms of interest correspond to redundant expansions and therefore a given signal can be represented by infinitely many coefficient vectors. To find a sparse vector among these is highly non-trivial (Candès et al. (2006)). Several properties of the transform vectors, e.g., their coherence determine whether this task can be performed in a computationally tractable manner. Recently, our understanding in this direction has improved by the explosion in compressed sensing literature. Outline of this thesis • First, we compare the construction and properties of four different transforms: the shift-invariant wavelet, the complex wavelet, the curvelet and the surfacelet transform. These all attempt to achieve a common goal a sparse representation of all signals in a given class (e.g., all seismic datasets or all natural images). These transforms have differ ent localization properties in the space and frequency domains, and as a result they all have unique and interesting properties. - • Second, we discuss how these transforms are tied into the theory of compressed sensing (Tsaig and Donoho (2006)). • And last, we investigate how these transforms perform in recovering signals from incomplete measurements (see e.g., Lustig et al. (2007)). In other words, we try to answer the question of which of the four transforms is most appropriate to use when recovering a signal from different types of measurements (either from the physical or the fre quency domains). Successful recovery of a signal (modeled as a vector in RN with poten tially very large N) from incomplete measurements depends mainly on three factors: (i) how sampling is done (e.g., uniform or non-uniform, in the spa tial domain or in the frequency domain), (ii) how compressible the signal is with respect to some prescribed transform (e.g., Fourier, wavelet, curvelet, etc.) and (iii) how mutually incoherent the measurement and sparsity do mains are. In recent years the interplay between these three key factors and their effect on successful sparse recovery is being investigated extensively. Recent results from compressive sampling give us a better understanding 2 Chapter 1. Introduction in this direction. However, these results do not necessarily apply to cases where we deal with structured and heavily redundant transforms such as shift-invariant wavelet, complex wavelet, curvelet or surfacelet transforms. 3 Chapter 2 The four transforms In this chapter we investigate the construction and properties of four dif ferent transforms. We start with two separable transforms shift-invariant and complex wavelets, and conclude with two non-separable transforms curvelets and surfacelets. - - 2.1 Motivation As mentioned in the previous chapter, many different transforms were de veloped for a wide variety of applications. In this work we focus on one particular type of data seismic wave fields. Some typical seismic wave fields are shown in Figure 2.1. Figure 2,1(a) shows a real seismic wavefield while Figure 2.1(b) shows a synthetic seismic wavefield. - (a) (b) Figure 2.1: Examples of seismic data sets. (a) real data (b) synthetic data. One can see that these datasets have a distinct structure. The wavefronts consist of hyperbolic and linear events, in some places with overlapping cusps. Furthermore, there are many singularities present in the data. An effective transform for this type of data should have the following useful properties: 4 Chapter 2. The four transforms Directional selectivity. Typical seismic data (shown in Figure 2.1) consists of wavefronts with a clear orientation. Thus the basis func tions of our transform should be able to align to these events. Sparsity. The transform should be able to capture most energy of the signal with only a few coefficients. Small redundancy. Real seismic datasets can be up to several giga bytes large, and even a moderate redundancy factor could significantly slow down the processing flow. Incoherence. As suggested by the theory of compressive sampling, the basis functions of the transform should be as incoherent as possible with the domain from which our measurement are acquired (Candès and Wakin (2007)). 2.2 Separable transforms A separable transform is constructed with separable products of a scaling function çb and a wavelet cL’. In this section we explore two transforms that are separable: shift-invariant wavelets and complex wavelets. 2.2.1 The shift-invariant wavelet transform Numerous variations of the wavelet transforms have been extensively stud ied and have appeared in literature as early as 1994 (Donoho and Johnstone (1994)), even though a mature mathematical theory of wavelets, as well as practical constructions of orthonormal wavelet bases was developed a decade later. Their ability to efficiently represent data that has multiscale struc ture has made the wavelet transform (WT) a powerful tool for researchers to use in all areas of signal processing. Perhaps the most widely implemented variation of the WT is the decimated bi-orthogonal wavelet transform (Ma sud and McCanny (1999)), which is utilized in the JPEG2000 compression algorithm (Dagher et al. (2003)). For other applications, such as edge detec tion and texture analysis, researchers continue to employ wavelet transforms with an arbitrary number of scales. This essentially means that the signal is modeled as a continuous function in some infinite dimensional function space, and one uses a wavelet basis or frame for this space. Because of computational constraints, this type of transform would only be practical for relatively small datasets. Usually these applications do not involve a reconstruction step. 5 Chapter 2. The four transforms In other applications, where a reconstruction from modified coefficients is required, researchers choose to use the undecimated wavelet transform (UDWT), which consists of a filter-bank construction but eliminates the decimation step in the orthogonal wavelet transform. These types of al gorithms are very fast and efficient in terms of computational complexity. Starck et al. (2007) have shown that utilizing an undecimated wavelet trans form for image denoising can lead to an improvement of up-to 2.5dB when compared to denoising via a decimated wavelet transform. The coefficients of the UDWT are computed using the same filter-bank as in the standard dec imated bi-orthogonal WT. This construction leads to three different bands of coefficients (vertical, horizontal and diagonal) for every scale. One-dimensional discrete wavelet transform The undecimated discrete wavelet transform UDWT (also called the shiftinvariant wavelet transform) was independently discovered by several au thors, see for example Shensa (1992). Next, we outline how the coefficients associated with UDWT are computed. We adopt the usual notation and denote the scaling and wavelet functions by and ?I’, respectively: qS(x/2) h[k]q(x = g[k]b(x (x/2) — — k), (2.1) k). (2.2) Here k e Z, x e R, and h[k} and g[k] are the analysis filters. Using the filter bank (h, g), the undecimated wavelet transform of a 1-D signal returns a set of coefficients W = {wi, wj, cj} where wj are the wavelet coefficients at scale j and cj are the coefficients at the coarsest scale. One can obtain the coefficients at the next finest scale from the ones at the previous scale using Mallat’s a trous algorithm (Holschneider et al. (1989), Shensa (1992), Starck et al. (2007)): ..., c+i[lj w[l] = = * (‘ * c)[l] cj)[l] = = > k], 3 h[k}ck[l + 2 (2.3) k], 3 g[k]ck[l + 2 (2.4) with — — f l h[l] 0 if 1/23 E Z, otherwise, (2 5) 6 Chapter 2. The four transforms where [n] h[—nj and [n] c[l] = g[—n]. The reconstruction is obtained by = * cj+i[l] + (2.6) * In the above equation, the filters Ii and refer to the synthesis of h and g, respectively. The filters are designed in such a way that the filter bank (h,g, h, ) satisfies the perfect reconstruction condition: )H(z) + G(z’)G(z) 1 H(z = 1, (2.7) where H, G, H and G are the z-transforms of h, g, h and respectively. Two dimensional discrete wavelet transform a trous The 2-D version of UDWT is obtained by extending the as follows: 1 1 C+l[k,1j [k,l] 1 w wi[k,1] [k,lj 1 w = h = (()3) = = Ii \1 1 algorithm *Cj)Lk,lj, 2.8 *c)[k,l], (2.9) *cj)[k,l], (2.10) *cj)[k,1], (2.11) where kg * c is the convolution of c by the separable filter kg, i.e., convolution first along the columns by h and then convolution along the rows by g (Starck et al. (2007)). The algorithm produces three wavelet images at each scale, ,w 1 w ,w 2 3 with each wavelet image being the same size of the original image. Consequently the redundancy factor of UDWT depends on the number of scales we choose to have in the decomposition, and is computed by is easy to compute; it is 3(L 1) + 1 where L is the number of scales. The key properties of the UDWT are: — 1. The UDWT is shift-invariant, i.e., a small shift in the signal does not perturb UDWT’s coefficient oscillation pattern around singularities. 2. In 2-D the UDWT has a redundancy factor of 3(L the number of scales in the decomposition. — 1) + 1 where L is It is important to note that although computational complexity of UDWT is O(LN), the redundancy factor of UDWT is exceedingly high for prac tical applications. If we are dealing with a moderately-sized dataset that we wish to analyze, perhaps with 6 scales, this already amounts to a redun dancy factor of 3(6—1)+1 = 16, which is unpractical for a lot of applications. 7 Chapter 2. 2.2.2 The four transforms The dual-tree complex wavelet transform (CWT) The complex dual-tree framework proposed by Selesnick et al. (2005) ex tends the work of Kingsbury (1999) by utilizing a filter bank implemen tation in conjunction with two real discrete wavelet transforms (DWTs). These DWTs use different sets of filters and satisfy the perfect reconstruc tion conditions, i.e., the analysis operation followed by synthesis operation would produce the original signal. One of the DWTs gives the real part of the CWT’s coefficients while the other DWT gives the imaginary part of the coefficients. If we let h 0 (n) and h 1 (n) be the lowpass/highpass pair for one of the filterbanks and go (n) and g (n) be the lowpass/highpass pair for the other 1 filter-bank than we can display the analysis/synthesis for the CWT in a schematic format, as shown in Figure 2.2. Figure 2.2: CWT filter bank (adapted from Selesnick et al. (2005)). Left: two levels of the analysis filter bank. Right: synthesis filter bank. These filters are designed in such a way so that the wavelet (t) := 9 (t), where /‘g (t) and /h (t) are real wavelet functions associated + i/’ with the two real wavelet transforms, are approximately analytic (see Lilly and Olhede (2008)). In this context, approximately analytic means that can be locally approximated by a convergent power series. The implementation of CWT is simple: the filter pair ho(n), hi(n) is used to implement one separable 2-D wavelet transform and the filter pair go(n), 91(n) is used to implement the other 2-D wavelet transform. Applying both transforms to the same 2-D image gives a total of 6 sub-bands: 2 HL, 2 LH and 2 HH sub-bands. The key properties of the CWT are outlined below; I)h (t) 1. The CWT is nearly shift invariant, i.e., a small shift in the signal does 8 Chapter 2. The four transforms not alter the coefficients’ oscillation pattern around singularities. 2. The CWT is a multiscale transform with six directionally-selective subbands in 2-D. 3. A complex wavelet is strictly localized in the spatial domain. 4. Regardless of the number of scales used in the decomposition, the redundancy factor of CWT is always 2 for n dimensions. 2.3 Non-separable transforms The shift-invariant and complex wavelet transforms both provide multiscale representations of signals. In addition, the CWT provides a limited number of directions in the representation. Next we explore two non-separable trans forms which provide more flexibility in choosing the number of directions at any given scale. 2.3.1 The curvelet transform From a conceptual point of view a curvelet (Candès and Donoho (2004)) is a scaled, shifted, and dilated copy of a basis element coa with width ap proximately equal to a and length approximately equal to A family of curvelets at scale a, location b, and orientation 0 are defined by coa,b,O(X) = — b)), (2.12) where R 0 is a rotation by 0 radians. In the continuous version of the curvelet transform, the parameter a E (0, 1), indicates the scale, 0 is the orientation, and b e R 2 is the spatial location. In the discrete version of the curvelet transform the parameters a and 0 are given by 3 a = 2’, and °j,l = 2irl 2_12j. . (2.13) At any given scale a curvelet is defined by applying translations and rotations to a “mother curvelet”. In the frequency domain it is defined by = 3 3 2 W 4 (23 / Il)V(2 2 0), (2.14) where W and V are smooth windows with compact support. Since W and V are compactly supported, one can see directly from Equation 2.14 that a curvelet will be strictly localized in the frequency domain. It turns out that 9 Chapter 2. The four transforms in the spatial domain curvelets are also well localized with fast decay. Once can see this idea demonstrated in Figure 4.16. The 2-D discrete curvelet transform, proposed by Candès and Donoho (2004), takes a 2-D signal of the form f(ni,n ), 0 2 < n as input and outputs a collection of coefficients that are obtained by computing inner products of the signal with curvelet window functions at different scales and directions (Candès et al. (2006)). The curvelet coefficients c(j, 1, k) at scale j, orientation 1 and spatial location k are defined by c(j,1,k) .pj1k(n1,n2)f(n1,n2), = (2.15) ‘1 ,fl2 where j,l,k (ni, n2) is the complex conjugate of the curvelet function j,1,k (ni, n2). The outline of the curvelet transform algorithm is as follows: 0 S A 2-D fast Fourier transform is applied to f(ni,n ) to get f(1,w2) with 2 —n/2 w, w2 < n/2. Next the Fourier samples of the input image are multiplied with the curvelet window functions Uj, (wi, ‘2) at different scales 1 and directions to obtain the product sj,l = 1 (w w U, , )f(w 2 ,w 1 ). This prod 2 uct is wrapped around the origin to produce W(s,j)(wi,w ). No wrapping 2 is required at coarsest and finest scales. Lastly, a windowed 2-D inverse fast Fourier transform is applied to 8 j,l, and the curvelet coefficients c(j, 1, k) are collected. The resulting collection of curvelet coefficients S a,O,& is a tight Parseval 0 Mallat (1999)) for Hilbert space: frame (see, e.g., a f = (f, SOa,o,i)SOa,o,i,, a,e,b WfW . 2 (f, SOa,6,b)1 = (2.16) a,8,b See Appendix A. 1 for the definition of frames for Hilbert spaces. Some of the important properties of curvelets are: 1. The curvelet transform provides a tight-frame expansion for any func (R as a series of curvelets. L ) tion f E 2 2. A single curvelet is strictly localized to one angular wedge in the frequency domain and has fast decay in the spatial domain. The effective support of Pj,l,k (x) obeys a parabolic scaling relation with j/2 . At scale j a curvelet is a ‘fat needle’ of length 2 2 width cc length and width 2i. 10 Chapter 2. The four transforms Figure 2.3: Frequency domain partitioning of the curvelet transform (adapted from Candès et al. (2006)). One curvelet is strictly localized to one angular wedge (shaded) in the frequency domain. 3, A curvelet is smooth along its major axis and is oscillatory along its minor axis (see Figure 4.16(a)). 4. The curvelet transform, as implemented in Candès et al. (2006), is approximately 8 times redundant in 2-D and 24 times redundant in 3D. 2.3.2 The surfacelet transform Curvelets provided us with a multiscale and a multidirectional signal de composition. We now explore surfacelet transforms, which are multiscale and multidirectional transforms that is less redundant than curvelets, and have a slightly different frequency partitioning. The surfacelet transforms, proposed by Lu and Do (2007), conceptually begins from a different construction than the curvelet transform. The sur facelet construction starts by extending the directional filter bank (DFB) that was proposed by Bamberger and Smith (1992) to an arbitrary num ber (N 2) of dimensions. The DFB partitions the frequency plane into 21,1 e N triangular wedges of equal area by radial lines passing through the origin of the (ui, w2) plane. In the case of a general N, the N-dimensional 11 Chapter 2, The four transforms filter bank (NDFB) partitions the frequency spectrum into rectangular pyra mids radiating from the origin. We consider a 3D example: to achieve the desired frequency partitioning, the NDFB (now N = 3) partitions the frequency cube into three hourglassshaped subbands that are aligned with the w, w2 and W3 axes, respectively. Next, two iteratively resampled checkerboard filterbanks (IRC) (Lu and Do (2007)), that act along the (ni, fl2) and (ni, n3) planes, are applied to each of the hourglass-shaped frequency subbands. The resulting output is 3 2 directional sabbands (Lu and Do (2006)). Next, we describe the construction of surfacelets, as give by (Lu and Do (2007): the Fourier transform of the signal gets inputed into first level of the multiscale pyramid where a highpass filter H(w) and a lowpass filter L(w) operate on the signal. This is similar to the CWT construction, where lowpass and highpass filters act on the signal to achieve a multiscale repre sentation (see Figure 2.2). The output of H(w) is inputed into the NDFB, and this returns the fine scale directional approximation to the signal. At the lowpass branch of the pyramid the signal is upsampled by a factor of two followed by an anti-aliasing filter 8(w). Finally, the signal is downsampled by a factor of 3. The output of this branch is a coarse approximation to the signal, which gets inputed into the next level of the multiscale pyramid. At the subsequent lowpass branches S(w) is no longer applied and the signal only gets downsampled by a factor of 2. Figure 2.4 describes the above process. Surfacelets were originally designed to get similar frequency partition ing as curvelets, but the two transforms aim to achieve this result with two very different approaches. The curvelet implementation starts with defining a”mother” curvelet in the Fourier domain. Moreover, its scaled and sheared copies form a partition of the frequency plane. As a result, the curvelet coefficients are obtained by multiplying the Fourier samples of the input signal with curvelet window functions at different scales and directions, fol lowed by spatial downsampling. The downsampling for curvelets is done in an alias-free fashion, which requires the curvelet functions to be strictly localized in the frequency domain. Since lower redundancy requires a higher subsampling ratio, this poses a trade-off between redundancy and spatial lo calization of curvelets, and subsequently implies a narrower transition band in the frequency supports. It is important to note that aliasing is allowed in the N-dimensional filter bank that is implemented in the surfacelet transform. At the end, the aliasing is cancelled out by carefully designed filters. As a consequence, 12 Chapter 2. The four transforms :4J NDFflH D () j)FB Ana’ysis Synthesis Figure 2.4: (Adapted from Lu and Do (2007)) Left: first two level of the analysis filter bank of the surfacelet transform. Subsequent levels of decomposition are achieved by recursively inserting the point a into afl+1. Right: synthesis filter bank of the surfacelet transform surfacelets are significantly less redundant than curvelets (a factor of 5 for surfacelets compared to a factor of 8 for curvelets). The surfacelet transform can use filters with fast spatial decay without loosing redundancy efficiency, since the filters do not need to be strictly bandlimmited (Lu and Do (2007)). Once again, we list the essential properties of the surfacelet transforms; 1. Just like the curvelet transform, the 2-D surfacelet transform provides a redundant, multiscale and multidirectional decomposition for func tions f. The transform is tight frame (see Appendix A.1). 2. Unlike curvelets, which are strictly localized to one angular wedge, a surfacelet is spread across an entire scale in the frequency domain. In the spatial domain surfacelets also obey a parabolic scaling relation of width o length , with most of the oscillations occurring along the 2 minor axis (see Figure 4.16). 3. In the 2-D case the surfacelet transform, as implemented in Lu and Do (2007), is approximately 5 times redundant. 4. Unlike the curvelet directional selectivity, where the number of wedges doubles at every other scale, in the surfacelet transform the number of wedges doubles at every scale. The finest and second-finest scales have the same number of wedges. 13 Chapter 3 Sparse seismic signal recovery 3.1 Motivation A seismic survey is an experiment to estimate properties of the Earth’s subsurface from reflected waves. The experiment usually consists of a source of energy, such as an air gun, or a vibrator (usually called a Vibroseis) and an array of receiver microphones that measure the reflected waves from the subsurface. By recording these acoustic pressure changes, it is possible to estimate the depth of the singularity that caused the reflection. Figure 3.1: A typical land seismic survey setup (adapted from www.iongeo.com). The survey consists of Vibroseis trucks and an array of geophones spread around the survey area. There are physical restrictions on where sensors can be placed: e.g., one cannot put sensors on the roads or the lake. As illustrated in Figure 3.1, seismic data is often irregularly sampled 14 Chapter 3. Sparse seismic signal recovery along the spatial coordinates due to practical constraints. The irregular sampling is usually due to the fact that at certain locations along the survey area a receiver physically cannot be placed. In marine surveys the irreg ular sampling occurs as a result of dead/severely contaminated traces or due to cable feathering. A lot of the multi-trace processing algorithms do not handle irregularly sampled data (data irregularities often result in postpro cessing image artifacts), and thus the data must be interpolated onto a regular grid. There exist relatively simple methods like linear interpolation of neighboring traces to handle this problem. Unfortunately this method does not account for wavefronts in the data, and the results of linear in terpolations are not always satisfactory. In practice, the methods that are used to interpolate seismic data are divided into three main groups: offset continuations (Stovas and Fomel (1993)), shot continuation (Schwab (1993)) and transform based methods (Sacchi et al. (1998)). In this work we build on the transform-based approach for signal recovery, as was proposed by Herrmann and Hennenfent (2008) and Hennenfent and Herrmann (2008b). Frequency domain methods for interpolation also require the data to be mapped onto a regular grid. This is often the case when the temporal fre quencies of a signal are sampled in an irregular manner (Lin et al. (2008)). This is an often-occurring issue in seismic wavefield modeling in the fre quency domain; when solutions to the time-harmonic Helmholtz equation are available for all frequencies, modeling in the frequency domain can be transformed back to modeling in the physical domain with time-stepping methods. However the solutions of the Helmholtz equation for each tem poral frequency are expensive to compute. On the other hand, if we are to invert a seismic wavefield from only a partial set of frequencies, aliasing artifacts will be visible in the resulting modeled wavefield, In this work we extend the ideas presented by Herrmann and Hennenfent (2008) and Hen nenfent and Herrmann (2008b) to interpolate the subsampled frequencies of a seismic signal to obtain an alias free result. 3.2 Model One application of transforms described in the previous chapter is recovering seismic wavefields from (i) subset of traces and (ii) a subset of angular frequencies. To describe the problem of signal recovery, we assume the forward model (3.1) y=RMf+n, 15 Chapter 3. Sparse seismic signal recovery where y is a vector, each component of which is a linear measurement of the full data vector f. The matrix M refers to the domain from which our measurements are acquired. In problem (i) M is the identity or the Dirac measurement basis, so M := I and in problem (ii) M is the Fourier measurement basis, so M := F. In this text we denote by ‘d the d x d identity matrix, and by Fd the d x d discrete Fourier transform matrix. We drop the subscript when it is not necessary to specify the dimensions. The matrix R is a restriction operator, i.e., it extracts those rows from M that represent the samples that are actually acquired. In real life problems the measurement basis and the restriction operator are determined by the acquisition geometry. In what follows, we wish to use some transform which “sparsifies” our signal. If we can find such a transform, we can then exploit this fact by approximating our signal by a linear combination of a few “x-let” atoms. More precisely, let S be the analysis operator of the “x-let” transform (whose atoms form a tight frame), and let SH, the conjugate transpose of S be the synthesis operator. Given a set of incomplete measurement y of our signal we wish to obtain an approximation f such that (i) f = SH* where * is sparse, (ii) {f f 112 is small. In certain cases, such an * can be computed by solving the optimization problem — * = argmin lixili subject to y — 2 <e. RMS”x11 (3.2) H refers to the synthesis matrix of one of the In what follows, the matrix 5 four transforms domains described in the previous chapter. This problem is setup in such a way that the solution has small L norm (which, at least under certain assumptions, ensures that it is sparse), while at the same time it fits the data up to a tolerance f (f is related to the amount of noise present in our data, and in the noise-free case = 0). Herrmann and Hennenfent (2008) have proposed to solve the problem given in Equation 3.2 when the underlying transform is the curvelet transform to recover a seismic signal from a subset of traces (Problem i). In this case in Equations 3.2 is the curvelet matrix, and M is the Dirac measurement basis. They use the principle of alignment, i.e., there is a large correlation between curvelets and wavefronts that locally have the same direction and frequency content, to explain heuristically why curvelets achieve good compressibility (one of the necessary conditions for signal recovery) of seismic data. 16 Chapter 3. Sparse seismic signal recovery 3.3 Compressive sampling In recent years, the theory of compressive sampling, pioneered by papers like Candès and Donoho (2004), Candès et al. (2006), and Donoho (2006) to name a few, has appeared in a lot of literature. As stated by Candès and Wakin (2007), data acquisition is usually extremely wasteful; large amounts of data are collected, only to be discarded later on. The acquired data is dis carded at the compression stage which is usually necessary for transmission and storing purposes. Compressive sampling works under the philosophy of as if “it were possible to directly acquire just the important information about the object of interest”. By taking only a small number of random projections, one has enough information to reconstruct the signal with an accuracy that is at least as good as the best compressed representation of the object. In this type of data acquisition, analog signals are essentially already in a compressed digital form. In principle, one can obtain “supper resolved” signals from just a few measurements and all that is required is to “decompress” the measured data. Candès and Romberg (2007) have shown that when our signal f is suffi ciently sparse, one can exactly recover the signal exactly via Li minimization. Theoremi (Candès and Romberg (2007)): Fix f E IR’ and suppose that the coefficient sequence x of f in the basis ‘I’ is S-sparse. Select m measurements in the domain uniformly at random. Then if m C . 2( ‘I’) 5 logn, . (3.3) . for some positive constant C and mutual coherence tion 3.2 is exact with overwhelming probability. j, the solution to Equa The quantity j is know as the coherence between the sensing matrix 1’ and the measurement matrix ‘1’, and is defined by (Candès and Romberg (2007)) ,Qi,’i)= max I(sok,bj)I, (3.4) 1k,3 fl where Yk is the kth row of and 1 bj is the j row of ‘I’. This quantity measures the spread of vectors from the measurement basis in the sparsifying domain. The “overwhelming probability”, mentioned in the above theorem ex 41) if m S. log(n/ö) (see Candès and Romberg ceeds 1 C (2007) for details). The theorem suggests a very specific type of acquisition; a non-adaptive sampling in an incoherent domain, followed by one-norms — . 17 Chapter 3. Sparse seismic signal recovery solution. Essentially, this would acquire the analog signal in a compressed form. After which, the data is “decompressed” via L minimization. The theory of compressive sampling does not really apply to our case as our “sparsity transforms” correspond to redundant expansions which are very structured and do possess bad coherence properties for any of these theorems to apply. Still, empirically L minimization seems to perform well to capture sparse representations in these cases. One of the main goals of this thesis is to speculate on possible empirical explanations for this unexpectedly well performance of compressed sensing-like methods, at least in the case of seismic datasets with the four mentioned transforms. 3.4 4 minimization algorithm The solution of the unconstrained optimization problem given in Equation 3.2 has been extensively studied and to this day many different solvers ex ist to tackle this problem. Without being exhaustive, a few of them are: iterative reweighed least-squares, introduced by Gersztenkorn et al. (1986), , proposed by Berg and Friedlander (2007), a variation of the pro 1 SPGL jected gradient method, proposed by Birgin et al. (2000), iterative hard thresholding, recently introduced by Blumensath and Davies (2008) , and iterative soft-thresholding, introduced by Daubechies et al. (2004). Herrmann and Hennenfent (2008) used the iterative soft thresholding (1ST) to solve the problem given in Equation 3.2. The 1ST is a very attractive solver to use, especially for large problems, since it only consists of matrix-vector multiplications and thresholding operations. However, before we can apply this solver directly to the constrained problem given in Equation 3.2, we must reformulate it as a series of uncon strained problems. Following the strategy of Elad et al. (2008), Problem 3.2 can be rewritten as an unconstrained problem using Lagrange multipliers A as RMSHx + AlIxili = argmin II (3.5) — where A is a parameter that controls the tradeoff between the sparsity of the solution (minimization of L norm) and the £2 data misfit. Now equation 3.5 is in such a form that we can directly apply the 1ST solver. To simplify the notation, we define the matrix A to be A := RMSH. Daubechies et al. (2004) have shown that iterative update of x by x —* T,(x + AT(y — Ay)) (3.6) 18 Chapter 3. Sparse seismic signal recovery converges to the solution of the problem give in Equation 3.5, for a particular value of )., as the number of iterations goes to infinity. In Equation 3.6, TA is a soft-thresholding function, defined by x—sign(x)A if IxI) iflxl<A. TA(x)—{ 0 (3.7) — The operation TA(x) thresholds and shrinks the entries of x by the threshold level ). An example of this function is shown in Figure 3.2 for ). = 3. C’) F x (a) Figure 3.2: Example of the soft-thresholding function TA(x) with a threshold level ) 3. To solve the full problem in Equation 3.5, the Lagrange multiplier ) must be cooled, and each subproblem is solved with this particular value of ).. Cooling is a common strategy to solve such problems (see e.g., Daubechies et al. (2004)); during this cooling process the subproblems are solved ap proximately for a decreasing A (Hennenfent (2008)). In practice, a limited number of inner iterations, about 5, are required to give an accurate esti mate for *A. In terms of computational complexity this is practical since 19 Chapter 3. Sparse seismic signal recovery each update of x in Equation 3.7 requires only two matrix-vector multipli cations. 1 2 3 4 5 6 7 Result: Estimate for x initialization; 0 x initial guess for x; Lagrange multiplier; initial )o L number of inner iterations; while IlAx Y112 E do for=1toLdo i—i+1; — — — — 8 ?k 5 ‘ 9 end 10 Ak+1 cxkAk (x + AH(y with 0 < — k < Axj)); 1 k—k+1; end Algorithm 1: Iterative soft thresholding, proposed by Figueiredo and Nowak (2003). 11 12 In the next chapter we apply the theory of compressive sampling com bined with £ minimization to investigate which transform performs best when recovering seismic signals from different types of restrictions. 20 Chapter 4 Results In this section, we demonstrate the performance of recovery using Algo rithm 1 on a synthetic seismic wavefield. We also compare the compressibil ity and mutual coherence of the transforms in the physical and frequency domains. We try to determine if there is indeed a correlation between the sparsity/coherence and the signal recovery performance of a particular trans form. 4.1 Recovering from a set of missing traces The first problem that we are considering is the sparse recovery of a seismic shot gather from a partial set of seismic traces. The model, consisting from a full set of traces and the data, consisting from a partial set of traces are shown in Figures 4.1(a) and 4.1(b) respectively. 100 Receiver position (m) 150 290 250 6 300 N C 100 Receiver position (m) 150 200 q d 250 390 . o 0 Full model in x—t domain (a) Data in x—f domain (b) Figure 4.1: Seismic data and missing traces. (a) Model synthetic seismic wavefield. (b) Data same seismic wavefield as in (a) but with 50% of the traces removed - - In this case the data is generated by randomly removing 50% of the traces. In section 4.2.2, we will demonstrate that the success of the recovery also 21 Chapter 4. Results depends on in which manner the data is sampled. We apply the iterative soft-thresholding algorithm to the data in Figure 4.1(b) and test the recov ery by the four different transforms curvelets, surfacelets, shift-invariant and complex wavelets. In this problem the measurement basis is the Dirac measurement basis. Mathematically, S in Equation 3.5 is one of the four transforms and M := I. - 100 Receiver position (m) 150 200 Receiver position Cm) 250 300 100 150 200 250 300 250 301 0 (0 0 NJ N Recovered Signal Recovered Signal (a) 100 Receiver position (m) 150 200 (b) 250 30 0 N 0 0 0 100 Receiver position (m) 150 290 C / Recovered Signal (c) Recovered Signal (d) Figure 4.2: Recovered seismic wavefield via (a) curvelets; (b) surfacelets; (c) shift-invariant wavelets; (d) complex wavelets The recovery results by the iterative soft-thresholding algorithm using curvelets, surfacelets shift-invariant wavelets and complex wavelets are plot ted in Figures 4.2(a), (b), (c) and (d) respectively. Using the 1ST algorithm (Algorithm 1), we have limited ourselves to 5 inner-loop and 30 outer-loop 22 Chapter 4. Results iterations. The signal-to-noise ratios, computed by SNR = —20 log 0 hf — 2/1og hIo 112, (4.1) where fo is the full model and f is the recovered signal, are: 7.49 dB for curvelets, 5.36 dB for surfacelets, 3.99 dB for shift-invariant wavelets and 4.34 dB for complex wavelets. As expected from the coherency argument and the empirical decay of coefficients, curvelets (the ones that are most incoherent with the Dirac measurement basis and have the fastest coefficient decay rate) are the most successful in the recovery. Compared to curvelets, shift-invariant and complex wavelets had very limited success. The recovery results from missing seismic traces are summarized in Table 4.1. Transform SNR (dB) Shift-invar. wavelet 3.99 complex wavelet 4.33 curvelet 7.49 surfacelet 5.36 Table 4.1: Physical restriction performance summary 4.2 Recovering from a set of missing angular frequencies In the second problem that we are investigating, we are assuming that our data comes from only a subset of temporal frequencies. The model and the data are shown in Figures 4.3(a) and 4.3(b), respectively. The data is generated by taking the frequency spectrum of the unrestricted synthetic wavefield from Figure 4.1(a) and randomly removing 50% of the temporal frequency content. In this case, just as in the previous case, the matrix S in Equation 3.5 is still one of the four transforms, but now the measurement basis is Fourier, so M := F. The problem of recovering a signal from a set of missing seismic traces (section 4.1) is an interpolation problem in the physical domain, while the recovery of a signal from a set of temporal frequencies is in fact an anti-aliasing problem in the physical domain. This idea is shown in Figure 4.4, where the data from Figure 4.3(b) is shown in the physical domain. The data contain aliasing artifacts that are the comparable to coherent noise. We again apply the iterative soft thresholding algorithm to test the recovery performance of the four different transforms. Again, just as in the previous problem, we limit ourselves to 5 inner-loop and 30 outer-loop iterations. 23 Chapter 4. Results (a) (b) Figure 4.3: Fourier spectrum and missing frequencies. (a)The model fre quency spectrum of the synthetic wavefield in Figure 4.1(a). (b) The data 50% of the temporal frequencies are missing. - - 100 Sample interval (1/rn) 150 200 250 300 0 0 I/ CD ‘p / \\ Figure 4.4: The data from Figure 4.3(b) in the physical domain. Aliasing artifact are present throughout the entire image. 24 Chapter 4. Results 190 Sample Interval (I/rn) 250 200 150 390 N 190 Sample Interval (I/rn) 150 250 200 390 NLI/ — Recovered Signol by curvelets — (b) (a) Sample interval (I/rn) 100 150 Recovered Signol by surfocelets 200 250 300 CM Sample interval (I/rn) 100150200250 300 c I Recovered Signol by shift—invorient wovelets (c) Recovered Signol by complex wovelets (d) Figure 4.5: Recovered seismic wavefield via (a) curvelets; (b) surfacelets; (c) shift-invariant wavelets; (d) complex wavelets. 25 Chapter 4. Results Sample Interval (1/rn) —n’ 0’ Sample interval (1/m) —2 0 0.2 9 (a) —0.4 Sample Interval (1/rn) —0.2 0 0.2 (c) (b) 0.4 — —0.4 Sample Interval (1/rn) —0.2 0 0.2 0.4 (d) Figure 4.6: Fourier spectra of the recovered signals via (a) curvelets; (b) surfacelets; (c) shift-invariant wavelets; (d) complex wavelets. 26 Chapter 4. Results The recovered signals by curvelets, surfacelets, shift-invariant wavelets and complex wavelets are shown in Figures 4.5(a), 4.5(c), 4.5(c) and 4.5(d). The corresponding Fourier spectra are shown in Figure 4.6. The signal-tonoise ratios for the four different cases are: 5.77 dB for curvelets, 6.17 dB for surfacelets, 11.0 dB for shift-invariant wavelets and 8.01 dB complex wavelets. Again, just as expected from the coherency argument, out of the four transforms shift-invariant wavelets had the best success in recovering the signal from missing frequencies. The wavelet recovery result has the highest SNR and the fewest aliasing artifacts left over. The recovery results from missing angular frequencies are summarized in Table 4.1. Transform SNR (dB) Shift-invar. wavelet 11.0 complex wavelet 8.01 curvelet 5.77 surfacelet 6.17 Table 4.2: Frequency restriction performance summary. 4.2.1 A different sampling scheme The sampling scheme used in section 4.2 assumes that we have no previous knowledge of the distribution of the frequency content of our data. That is, we assume the the frequencies are uniformly distributed over the full spectrum, and we sample this spectrum in a uniform manner. From Figure 4.3(b) it can be seen that most of the frequency content of the wavefield is concentrated around the 5 Hz to 50 Hz band. If we have this previous knowledge we can use this to sample our data only in the region that is know to have to highest frequency content. The averaged frequency spectrum, shown in Figure 4.7(a) is now used as a pseudo probability density function to determine where our samples will come from. That is, the averaged spectrum is used as an envelope for the new sampling scheme, i.e., most of the samples are measured from frequency intervals that have the highest relative magnitude. The signal in Figure 4.7(b) shows the recovered result by taking the samples from Figure 4.7(a) (red lines). As expected, we have a signifi cant improvement over the purely random sampling scheme result in Figure 4.5(c). The signal-to-noise ratio for the band random sampling scheme is 11.86 dB, while in the uniform random sampling case the SNR is 11.0 dB. This is a significant improvement since in this case we were recovering from only 20% of the original temporal frequencies while in the previous case we were recovering from 50% of the original frequencies. 27 Chapter 4. Results S (a) 100 150 Trace # 200 250 300 d C çd V Recovered Signal by shift—invarien+ wavelets (b) Figure 4.7: Frequency-weighted sampling scheme. (a) Averaged frequency distribution with 20% of the frequencies used to sample the signal (red lines); (b) recovered signal via shift-invariant wavelets with the frequency samples taken from (a). 28 Chapter 4. Results 4.2.2 The importance of random sampling Lets return to the scenario where we assume to have no previous knowledge on the location of the frequency content of the data. One might think that if this is the case, then instead of sampling the data randomly, where we could miss locations of high frequency content, we could try to uniformly sample the data. In this part we will show that the success of signal recovery is very much dependent on the sampling scheme used. Lets first consider a id example. signa’ spectra Figure 4.8: A 1D signal and it’s Fourier spectrum The signal in Figure 4.8 is a signal composed of a superposition of three cosine functions with different frequencies. The corresponding Fourier spec trum has three distinct peeks corresponding to the three different frequen cies. Lets subsample the signal in a random manner (in the physical do main). 29 Chapter 4. Results randorrIy sarpIed signal spectra Figure 4.9: Randomly subsampled 1D signal and it’s Fourier spectrum Figure 4.9 show the original signal randomly subsampled by a factor of 3. The red lines represent the samples actually acquired. The Fourier spectra of the subsampled signal is noisy, but never-the-less we can still see the three distinct spikes corresponding to the three different frequencies. If we set to zero all the coefficients that are below the threshold level (the green line), that when we synthesize back to the physical domain, we can exactly recover our signal. Now lets subsample the signal in a uniform manner. 30 Chapter 4. Results spectra Figure 4.10: Uniformly subsampled 1D signal and it’s Fourier spectrum The signal in Figure 4.10 is a three-fold uniformly sabsampled signal. The Fourier spectrum is no-longer noisy; it now consists of the three orig inal spikes plus some coherent aliasing artifacts. All of these spikes live above the threshold level (that same threshold level as in the previous Fig ure). Therefore, after thresholding and synthesizing back to the physical domain we would not get back our original signal, but instead a distorted and aliased copy of it. We now demonstrate this idea on our 2-D synthetic wavefield. Figure 4.11(c) shows a three-fold uniformly subsampled frequency spectrum of the synthetic wavefield. Figure 4.11(b) shows the attempt to recover the signal with the iterative soft-thresholding algorithm using the shift-invariant wavelet transform. Clearly the algorithm is unsuccessful since we have co herent aliasing artifacts left in the recovered result. In this case the signal to noise ratio is 1.58 dB. 4.2.3 Direct Fourier spectra interpolation Motivated by the success of curvelets to recover a signal from a set of missing seismic traces, we would like to see if we can improve upon the recovery of a signal from a set of missing angular frequencies, using curvelets. This is done by direct interpolation of the subsampled Fourier spectra. In this case 31 Chapter 4, Results (b) (a) (c) Figure 4.11: Uniform sub-sampling of angular frequencies. (a) Three-fold uniformly subsampled frequency spectra of the synthetic wavefield; (b) At tempted recovery result by shift invariant wavelets with a uniform subsam pling scheme from (a); (c) interpolated spectrum. 32 Chapter 4. Results the optimization problem can be formulated as * = argmin WxIi subject to 2 I’ Ax11 (4.2) — 0 and SH is the synthesis of the curvelet where A = RISH, y = FRIf transform. The approximation to f 0 is calculated by f FHSH*. —0.4 Wavenumber (1/rn) 0.2 —0.2 9 Sample Interval (m) 150 200 100 0.4 250 300 * 0 \\\ 0 7/ (a) (b) Figure 4.12: Subsampled signal. (a) Subsampled Fourier spectrum of a seismic signal; (b) same signal as in (a), but in the physical domain. The subsampled Fourier spectrum and the signal in the physical domain are shown in Figure 4.12. The full model is shown in Figure 4.12(b). The spectra in Figure 4.12(a) is two-fold under-sampled. 190 Sample interval (rn) 150 200 250 390 0 / 0 -. / (a) (b) Figure 4.13: Directly interpolating the Fourier spectra. (a) Interpolated spectra from Figure 4.12(a); (b) interpolated signal in the physical domain. 33 Chapter 4. Results The results of direct Fourier spectra interpolation are shown in Figure 4.13. The interpolated spectrum is shown in Figure 4.13(a) and the recov ered signal in the physical domain is shown in Figure 4.13(b). Again the 1ST solver, with 5 inner and 30 outer ioop iterations was applied to Problem 4.2 to obtain these results. Unfortunately the results from directly interpolation the Fourier spec trum are not as favorable as the ones observed in section 4.2. The recovered signal in Figure 4.13(b) has a significant number of aliasing artifacts left over. The SNR for the recovered signal is 4.64 dB. Using shift-invariant wavelets we observed better results when recovering a signal from missing angular frequencies, as formulated in Problem 3.2 The wavelet recovery for a two-fold subsampled signal in the frequency domain had a SNR of 11.0 dB. 4.3 Compressibility In this section, we compare the compressibility of seismic data with respect to the four transforms. In this context, compressibility means that only a a few coefficients are required to capture most of the energy of a signal. 4.3.1 Nonlinear approximation Since each of the four transforms provides an over-complete signal repre sentation with different redundancy factors, it would not be a fair to just compare the decay rates of the sorted coefficients corresponding to each dif ferent transform. Two different approaches have been used in the literature to remedy the problem of different redundancy factors (see e.g., Tantum and Collins (2003)). The first approach is a Monte Carlo type of sampling of coefficients where we account for the different redundancies. For exam ple, the curvelet transforms is approximately 8 times redundant (in 2-D) so we randomly remove 7 out of 8 of the curvelet coefficients. This pro cedure is repeated n times (n should be large, say 100) and the resulting coefficients are averaged and the resulting averaged coefficients are plotted in magnitude-descending order. The same procedure is done for the other three transforms. The results are shown in Figure 4.14(a). The other approach is, instead of analyzing at the relative decay rate of the coefficients we can investigate the mean squared error (or the £2 error) as a function of the percentage of coefficients used in the reconstruction. This type of comparison is shown in Figure 4.14(b). 34 Chapter 4. Results From the two plots in Figure 4.14 we can see that curvelet coefficients decay in magnitude significantly faster than the other three transforms’ co efficients. The results are similar in the compression plot a significantly smaller fraction of coefficients are required to attain the same £2 error from a partial reconstruction of curvelet coefficients than from any other transforms partial set of coefficients, i.e., 30% of shift-invariant wavelet coefficients are required to get the same MSE compared to when we use only 16% of curvelet coefficients. In both plots (compression and £2 error) the experiments were performed on a seismic dataset, shown in Figure 4.1(a). The same number of scales (5) was used in every transform. From Figures 4.14(a) and 4.14(b) we get an idea about the compressibil ity of seismic datasets with respect to different transforms. But never the less this is a rather rudimentary way of trying to determine which transform will perform better in seismic signal recovery. The Pareto curve (discussed in the next section) is perhaps a more suitable tool. - 35 Chapter 4. Results a) D (a 2 G) > (a (a) CI) 10 15 20 25 30 % of coefficients used in reconstruction (b) Figure 4.14: Nonlinear approximation. (a): Randomly sampled transform coefficients plotted in magnitude-descending order. (b) MSE of the recon struction from an incomplete set of transform coefficients 36 Chapter 4. Results 4.3.2 The Pareto curve Another tool to measure the compressibility of a signal could be obtained us ing the Pareto curve (see e.g., Berg and Friedlander (2008)) associated with the optimization problem given in Equation 3.2, The use of this curve has been previously proposed to compare the compressibility of seismic signals by curvelets and Fourier in Hennenfent and Herrmann (2008a). We extend this approach to include surfacelets, shift-invariant and complex wavelets in the compressibility comparison, and test whether one can indeed infer the efficiency of a transform for sparse recovery from the corresponding Pareto curves. The Pareto curve traces the optimal tradeoff between the data misfit and some prior. We use SPG 1 (Berg and Friedlander (2007)) to solve the basis pursuit problem (BP ) (Chen et al. (1999)), where BP is 0 BP : mm lxiii s.t. iy — 2 <o. Axii (4.3) In our case we are assuming a noise-free dataset ( = 0). The matrix A corresponds to the synthesis matrix of one of the above-described transforms and y is the data. The columns of each A are normalized to have unit-norm. Stated in plain English, we try to find a vector of coefficients that has the smallest £ -norm that fits our data. 1 The lines in Figure 4.15 represent the Pareto curves for the four different transforms, plus Fourier. The curves were computed on a synthetic seismic dataset shown in Figure 4.1(a). For each transform, the point at the bottom of the ii Ax11 0 2 axis represents the £ norm of the solution to the BP problem, i.e., the vector of transform coefficients with the smallest L norm that perfectly describes the data. In Figure 4.15 we can see that Fourier achieves the poorest compressibil ity of the signal. Curvelets provide the sparsest representation of the data with iiXcurvelet iii = 32.6, the smallest L norm solution among all transforms. Hennenfent and Herrmann (2008a) suggested that this type of analysis pro vides a better understanding of signal compressibility than just comparing the empirical decay rates of the magnitude sorted coefficients as was done by Candès et al. (2006). By tracing the solution path to the BP problem the Pareto curve provides us with a concrete way of comparing transform compressibility. The compressibility comparison of the magnitude-sorted co efficients is more suitable when comparing transforms which have the same redundancy factors. It is important to note that the transforms’ recovery performance, observed in the next chapter, does not correspond to either — 37 Chapter 4. Results >c 40 II.iI Figure 4.15: Pareto curves for the different transforms. The lined in the plot represent Pareto curves for compressibility of seismic data by: Fourier (yellow), curvelets (blue), surfacelets (red), complex wavelets (green) and shift-invariant wavelets (black) the order of BP solutions observed in the Pareto curve, nor the empirical decay rates of the sorted coefficients. 4.4 Mutual coherence In this section we adapt this theory of compressive sampling to our prob lems of interest, i.e., we would like to determine the coherence i of the two measurement domains Dirac (M := I) and Fourier (M := F), for the four transform domains curvelets, surfacelets, shift-invariant and complex wavelets. Although it is computationally expensive to calculate this quan tity even for small 2-D signals, we can get an idea of its relative values from - 38 Chapter 4. Results one ‘x-let’ in the physical domain and its corresponding frequency response. Figure 4.16 demonstrates this coherence idea. It shows one curvelet, sur facelet, shift-invariant wavelet and complex wavelet plotted in the physical and frequency domains, all at the same scale. One can see that a curvelet appears to be most incoherent in the physical domain (i.e. (I, Scurveiet) S the smallest), that is, there are large oscillations along one direction, and it appears to be the least-localized in the physical domain. Shift-invariant wavelets appear to be the least localized in frequency and thus they are most incoherent in the Fourier domain ( (F, Swaveiet) is the smallest). The basis functions in Figure 4.16(a) were generated by applying a syn thesis transform on a vector with a single non-zero entry. The corresponding frequency response, Figure 4.16(b), is generated by taking an fft of the basis functions in Figure 4.16(a), applying an fftshift (zero-frequencies are now in the center) and plotting their absolute values. 4.4.1 Coherence approximation A more deterministic way of looking at the coherence of the four transforms is to use a Monte-Carlo (Nagata and Watanabe (2008)) type of method. Rather than calculate the full set of inner products, as defined by Equation 3.4, we approximate the coherence for only a subset of columns as follows; start by defining the quantity vj by z’j = max IATAviI with A = RMWSH, (4.4) where S is one of the four sparsifying transforms, W is a diagonal matrix that ensures that the columns of A have unit norm, M is am measurement basis (either physical or frequency) and R is a restriction operator. In the physical domain R restricts the columns that correspond to missing traces and in the frequency domain R restricts the rows that correspond to missing angular frequencies. The vector v is a vector that singles out one column of A, i.e., v is a vector of zeros with a single 1 at a random location. The the coherence of the transform and the measurement basis is then computed by /2 = max{v}E[1N], (4.5) where vj is defined in Equation 4.4. N is the number of columns for which we wish to calculate /2. We perform this computation on 10000 random vectors v and the results are summarized in Table 4.3. 39 Chapter 4. Results Curvelets Surfacelets Complex Wavelets Shift-invar. wavelets Physical Domain jEt mean var 0.876 0.748 0.0056 0.871 0.692 0.0042 0.998 0.738 0.0305 0.997 0.741 0.0413 Frequency Domain mean var 0.924 0.818 0.0010 0.841 0.691 0.0037 0.876 0.612 0.0499 0.810 0.690 0.0091 Table 4.3: Coherence approximations The mean and variance are calculated from the set {v} for ii as defined in Equation 4.4. From the coherence approximations in Table 4.3, one can see that complex and shift-invariant wavelets are the most coherent in the physical domain. Shift-invariant wavelets are the most incoherent with the frequency domain. One could use this Table 4.3 as a rough-predictor of which transform will be the most suitable for signal recovery in the two different cases. Curvelets are the most coherent in the frequency domain, and are quite incoherent in the physical domain so we expect curvelets to perform better when recovering a signal from missing traces rather than when we recover a signal from missing frequencies. By the same rational we can expect the best results when recovering a signal from missing frequencies using shift-invariant wavelets. Although we observed the best result when recovering a signal from missing traces (see Chapter 4.1) using curvelets, surfacelets appear to be slightly more incoherent in the physical domain than curvelets. 4.4.2 Data-dependent coherence approximation We hypothesize that the success of the recovery of a signal in the physi cal domain using curvelets might be due to the nature of the signal i.e. curvelets achieve the sparsest representation of seismic signals (see Section 4.3.1). Instead of randomly choosing the v vectors in Equation 4.4, we choose the vectors that correspond to 10000 most significant or insignificant transforms’ coefficients, and repeat the same calculation for l. - 40 Chapter 4. Results Curvelets Surfacelets Complex Wavelets Shift-invar. wavelets Physical Domain mean var 0.864 0.825 0.0013 0.834 0.677 0.0040 0.875 0.556 0.0298 0.871 0.599 0.0301 Frequency Domain mean var 0.864 0.819 0.0007 0.833 0.681 0.0035 0.875 0.575 0.0316 0.872 0.582 0.0211 Table 4.4: Data-dependent coherence approximations using 10000 most sig nificant transforms’ coefficients Curvelets Surfacelets Complex Wavelets Shift-invar. wavelets Physical Domain mean i var 0.845 0.816 0.0073 0.833 0.698 0.0031 0.875 0.875 1.4e-26 0.871 0.871 6.4e-19 Frequency Domain mean var 0.846 0.815 0.0082 0.834 0.671 0,0029 0.875 0.666 0.0490 0.822 0.670 0.0517 Table 4.5: Data-dependent coherence approximations using 10000 most in significant transforms’ coefficients From Tables 4.4 and 4.5 no clear pattern can be observed that would indicate which transform should perform better when it comes to recovering a signal from either missing traces or missing angular frequencies. That is, both tables shows that surfacelets are slightly less coherent than curvelets in the physical domain, but when recovering a signal from missing traces (phys ical restriction), the best results are observed using curvelets (see Section 5). 41 Chapter 4. Results (a) (b) Figure 4.16: Single atoms of the four transforms in the physical and fre quency domains. (a): curvelet (top right), surfacelet (top right), shift invariant wavelet (bottom left), complex wavelet (bottom right) in the phys ical domain. (b): Fourier transform of the atoms in the same order. 42 Chapter 5 Conclusions and discussion Throughout this thesis, we have investigated the recovery of a seismic sig nal from missing angular frequencies and missing seismic traces using four different transforms: shift-invariant wavelets, complex wavelets, curvelets and surfacelets. The recovery was carried out using the iterative soft thresholding procedure, and the algorithm was ran for a fixed number of iterations. The signal-to-noise ratios for the different experiments performed are summarized in Table 5.1. Recovery signal-to-noise ratios Physical restriction Frequency restriction 50% of traces missing 50% of frequencies missing Curvelets 7.49 dB 5.77 dB Surfacelets 6.17dB 5.36 dB Shift-invar. wavelets 3.99 dB 11.0 dB Complex wavelets 4.33 dB 8.01 dB Table 5.1: Recovery performance summary In the problem of recovering the signal from missing traces the best results were observed using the curvelet transform. Curvelets were able to give the sparsest representation of or seismic signal at the cost of a moderate redundancy factor. Also, out of the four transforms curvelets appear to be most incoherent with the Dirac measurement basis, which explains, from a compressive sampling point of view why the best interpolation results were observed using curvelets. In the problem of recovering a signal from a subset of angular frequencies the best results were observed using the shift-invariant wavelet transform, although at a cost of high redundancy. These results were found to be significantly better (an improvement by 6.36 dB) than a direct interpolation of the Fourier spectra of the subsampled signal. Out of the four transforms, shift invariant wavelets were observed to be most incoherent with the Fourier measurement basis. We have also seen that the success of the recovery also depends on the 43 Chapter 5. Conclusions and discussion sampling scheme used. Uniform subsampling results in coherent aliasing artifacts and as a result the recovery algorithm fails to produce meaningful results, as can bee seen in Figure 4.11(b). On the other hand, if one has some prior knowledge of where most of the frequency content is concentrated, one can sample in the band. In this case the recovery results had a high SNR (11.86 dB) when recovering from just 20% of the frequencies. Throughout this thesis we have attempted to find a direct link between a transform’s performance in recovering seismic signals from incomplete mea surement and the compressibility/mutual coherence of a transform with a particular measurement basis. Unfortunately no such link was observed. This is mainly due to the fact that our sparsifying transforms correspond to redundant expansions with poor coherence properties. A more thorough in vestigation is required to determine precisely how non-orthonormal systems (like shift-invarieant wavelets, complex wavelets, curvelets and surfacelets) tie into the theory of compressive sampling. 44 Appendix A Wavelet concepts A.1 Frames Frame theory is used to analyze the completeness, stability and redundancy of discrete transforms. It was originally developed by Duffin and Schaefer (1952) to reconstruct band-limited signals f from irregularly spaced samples in a {f(t)}z. Roughly speaking, a frame is a family of vectors Hilbert space H that characterizes any signal f e H from its inner prod ucts { (f, g) }nEF in a numerically stable way. More precisely, the sequence {n}nEr E H is a frame of H if there exist two constants A > 0 and B > 0 such that for any f E H AIIf 12 nr )I m (f 2 Bfj{. (A.1) When A = B the frame is said to be tight. A frame defines a complete and stable signal representation, which may or may not be redundant. The redundancy of the signal representation is reflected by the frame bounds A and B if the frame vectors have unit norm. Indeed, a frame {çO}, with = 1, Vn e F, is an orthonormal basis if and only if the frame bounds satisfy A = B = 1. 45 Bibliography Bamberger, R. H. and M. J. T. Smith, 1992, A filter bank for the directional decomposition of images: theory and design: IEEE Transactions on Image processing, 40, 882—893. P. Friedlander, Berg, E. v. and M. A solver for large-scale sparse (http://www.cs.ubc.ca/labs/scl/spgll/?n=HomePage). 2007, Spgll: reconstruction. 2008, Probing the pareto frontier for basis pursuit solutions: Tech. Rep. TR-2008-01, Department of Computer Science, University of British Columbia, Vancouver. (To appear in SIAM J. Sci. Comp.). Birgin, E. G., J. M. MartInez, and M, Raydan, 2000, Nonmonotone spectral projected gradient methods on convex sets: SIAM J. on Optimization, 10, 1196—1211. Blumensath, T. and M. E. Davies, 2008, Iterative hard thresholding for compressed sensing (technical report). Candès, E. and J. Romberg, 2007, Sparsity and incoherence in compressive sampling: Inverse Problems, 23, 969—985. Candès, E., J. Romberg, and T. Tao, 2006, Robust uncertainty principles: exact signal reconstruction from highly incomplete frequency information: Information Theory, IEEE Transactions on, 52, 489—509. Candès, E. J. and L. Demanet, 2005, The curvelet representation of wave propagators is optimally sparse: Communications on Pure and Applied Mathematics, 58, no. 11, 1472—1528. Candès, E. J., L. Demanet, D. L. Donoho, and L. Ying, 2006, Fast discrete curvelet transforms: Multiscale Modeling and Simulation, 5, no. 3, 861— 899. Candès, E. J. and D. L. Donoho, 2004, New tight frames of curvelets and optimal representations of objects with piecewise-C 2 singularities: Com munications on Pure and Applied Mathematics, 57, no. 2, 219—266. 46 Bibliography Candès, E. J. and M. Wakin, 2007, People hearing without listening: an introduction to compressive sampling.: To appear in IEEE Transactions on Signal Processing. Chen, 5. 5,, D. L. Donoho, and M. A. Saunders, 1999, Atomic decomposi tion by basis pursuit: SIAM Journal on Scientific Computing, 20, 33—61. Dagher, J., A. Bilgin, and M. Marcellin, 2003, Resource-constrained rate control for motion jpeg2000: Image Processing, IEEE Transactions on, 12, 1522—1529. Daubechies, I., M. Defrise, and C. De Mol, 2004, An iterative thresholding algorithm for linear inverse problems with a sparsity constraint: Commu nications on Pure and Applied Mathematics, 57, no. 11, 1413—1457. Donoho, D., 2006, Compressed sensing: Information Theory, IEEE Trans actions on, 52, 1289—1306, Donoho, D. L. and I. M. Johnstone, 1994, Ideal spatial adaptation by wavelet shrinkage: Biometrika, 81, 425—455. Duffin, R. J. and A. C. Schaefer, 1952, A class of nonharmonic fourier series: Trans. Amer. Math. Soc., 72, 341—366. Elad, M., J.-L. Starck, P. Querre, and D. L. Donoho, 2008, Simulataneous cartoon and texture image inpainting using morphological component anal ysis (mca): 19, 340—358. Figueiredo, M. and R. Nowak, 2003, An em algorithm for wavelet-based image restoration: Image Processing, IEEE Transactions on, 12. Gersztenkorn, A., J. B. Bednar, and L. R. Lines, 1986, Robust iterative inversion for the one-dimensional acoustic wave equation: Geophysics, 51, 357—368. Hennenfent, G., 2008, Sampling and reconstruction of seismic wavefields in the curvelet domain: PhD thesis, The University of British Columbia, Vancouver, BC Canada. Hennenfent, G. and F. J. Herrmann, 2008a, One-norm regularized inver sion: learning from the pareto curve: Submitted to the SEG 2008 conven tion. 2008b, Simply denoise: wavefield reconstruction via jittered under sampling: Geophysics, 73, no. 3. Herrmann, F. J. and G. Hennenfent, 2008, Non-parametric seismic data recovery with curvelet frames: Geophysical Journal International, 173, 233—248. 47 Bibliography Holschneider, M., R. Kronland-Martinet, J. Monet, and P. Tchamitchian, 1989, Wavelets, time-frequency methods and phase space: Springer Verlag,Berlin. Kingsbury, N., 1999, Image processing with complex wavelets: Philosoph ical Transactions: Mathematical, Physical and Engineering Sciences, 357, 2543—2560. Lilly, J. M. and S. C. Olhede, 2008, On the design of optimal analytic wavelets. (techical report). Lin, T. Y., E. Lebed, Y. A. Erlangga, and F. J. Herrmann, 2008, Interpolat ing solutions of the helmholtz equation with compressed sensing: Presented at the SEG. Lu, Y. and M. N. Do, 2006, Video processing using the 3-dimensional surfacelet transform: Signals, Systems and Computers, 2006. ACSSC ‘06. Fortieth Asilomar Conference on, 883—887. Lu, Y. M. and M. N. Do, 2007, Multidimensional directional filter banks and surfacelets: Transactions on Image Processing, 16, no. 4, 918—931. Lustig, M., D. Donoho, and J. M. Pauly, 2007, Sparse mn: The applica tion of compressed sensing for rapid mr imaging: Magnetic Resonance in Medicine, 8, 1176—1194. Mallat, 5., 1999, A wavelet tour of signal processing, second edition: Aca demic Press. Masud, S. and J. McCanny, 1999, Rapid vlsi design of biorthogonal wavelet transform cores: Signal Processing Systems, 1999. SiPS 99. 1999 IEEE Workshopk, 291—300. Nagata, K. and S. Watanabe, 2008, Exchange monte carlo sampling from bayesian posterior for singular learning machines: Neural Networks, IEEE Transactions on, 19, 1253—1266. Sacchi, M. D., T. J. Ulrych, and C. J. Walker, 1998, Interpolation and extrapolation using a high-resolution discrete Fourier transform: Transac tions on Signal Processing, 46, no. 1, 31—38. Schwab, M., 1993, Shot gather continuation. (technical report). Selesnick, I. W., R. G. Baraniuk, and N. G. Kingsbury, 2005, The dual tree complex wavelet transform: IEEE Transactions, Signal processing, 123—15 1. 48 Bibliography Shensa, M., 1992, The discrete wavelet transform: Wedding the a trous and mallat algorithms: IEEE Transactions on Information Theory, 40, 2464—2482. Starck, J.-L., J. Fadili, and F. Murtagh, 2007, The undecimated wavelet decomposition and its reconstruction.: IEEE Transactions on Image Pro cessing, 16, 297—309. Stovas, A. M. and S. B. Fomel, 1993, Kinematically equivalent dmo oper ators: Russian Geology and Geophysics, 37, 102—113. Tantum, S. and L. Collins, 2003, Performance bounds and a parameter transformation for decay rate estimation: Geoscience and Remote Sensing, IEEE Transactions on, 41, 2224—2231. Tsaig, Y. and D. L. Donoho, 2006, Extensions of compressed sensing: Sig nal Process., 86, 549—571. 49
- Library Home /
- Search Collections /
- Open Collections /
- Browse Collections /
- UBC Theses and Dissertations /
- Sparse signal recovery in a transform domain
Open Collections
UBC Theses and Dissertations
Featured Collection
UBC Theses and Dissertations
Sparse signal recovery in a transform domain Lebed, Evgeniy 2008
pdf
Page Metadata
Item Metadata
Title | Sparse signal recovery in a transform domain |
Creator |
Lebed, Evgeniy |
Publisher | University of British Columbia |
Date Issued | 2008 |
Description | The ability to efficiently and sparsely represent seismic data is becoming an increasingly important problem in geophysics. Over the last thirty years many transforms such as wavelets, curvelets, contourlets, surfacelets, shearlets, and many other types of ‘x-lets’ have been developed. Such transform were leveraged to resolve this issue of sparse representations. In this work we compare the properties of four of these commonly used transforms, namely the shift-invariant wavelets, complex wavelets, curvelets and surfacelets. We also explore the performance of these transforms for the problem of recovering seismic wavefields from incomplete measurements. |
Extent | 1390514 bytes |
Subject |
Wavelets Transforms Sparsity |
Genre |
Thesis/Dissertation |
Type |
Text |
FileFormat | application/pdf |
Language | eng |
Date Available | 2009-02-04 |
Provider | Vancouver : University of British Columbia Library |
Rights | Attribution-NonCommercial-NoDerivatives 4.0 International |
DOI | 10.14288/1.0070795 |
URI | http://hdl.handle.net/2429/4171 |
Degree |
Master of Science - MSc |
Program |
Mathematics |
Affiliation |
Science, Faculty of Mathematics, Department of |
Degree Grantor | University of British Columbia |
GraduationDate | 2008-11 |
Campus |
UBCV |
Scholarly Level | Graduate |
Rights URI | http://creativecommons.org/licenses/by-nc-nd/4.0/ |
AggregatedSourceRepository | DSpace |
Download
- Media
- 24-ubc_2008_fall_lebed_evgeniy.pdf [ 1.33MB ]
- Metadata
- JSON: 24-1.0070795.json
- JSON-LD: 24-1.0070795-ld.json
- RDF/XML (Pretty): 24-1.0070795-rdf.xml
- RDF/JSON: 24-1.0070795-rdf.json
- Turtle: 24-1.0070795-turtle.txt
- N-Triples: 24-1.0070795-rdf-ntriples.txt
- Original Record: 24-1.0070795-source.json
- Full Text
- 24-1.0070795-fulltext.txt
- Citation
- 24-1.0070795.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.24.1-0070795/manifest