Open Collections

A model-based approach to complex contour generation for process automation using computer vision

University of British Columbia

The research described in this thesis addresses the development of methodology for fast extraction of complex processing information of an object using relatively simple on-line measurements and prior knowledge and information that can be generated off-line on the object. Even though the approaches investigated here are quite general, the techniques are developed with respect to the specific application of flexible automation of a fish processing cell. Robotics and computer vision have been combined to produce useful results in a variety of applications. Vision guidance may be applied at least in two ways here. Cutting contours of fish could be generated off-line and these could be used to generate the reference signals for the cutting controller. Alternatively, cameras mounted on the cutter could be used to guide the cutter in real time. The present research concentrates on the first approach since the objective is to generate the cutting contour sufficiently fast to meet the process speed requirements. The complexity of the vision problem is eased to some extent by the fact that the object (fish) is already recognized, at least in a generic sense. Low-to-medium-level computer vision techniques involving image transformation, enhancement and analysis of basic features such as edges, regions, shape, colour and texture, are reasonably well established and widely applied. The application of these techniques to directly generate the cutting contour of an arbitrary fish would be computationally slow and unacceptable in terms of process speed requirements. The present research effort is directed at expediting vision-based generation of cutting contours through the use of model-based vision techniques. One of the goals of this research is to develop a knowledge-base and a model-base to support the complex feature extraction procedure. Use of low-level image analysis techniques to obtain non-complex features at high speeds from the fish images is a further goal, as this is related to the first goal. In the model-based approach, the fish models are generated using a representative database of measurements on fish. The data includes a cutting contour and some dimensional measurements which are referred to as attributes or features. The cutting contours are non-dimensionalized, transformed, and grouped into an appropriate number of models using a systematic grouping procedure. Each model contains a non dimensional cutting contour for the corresponding group of fish, and a set of attributes. In real-time operation, the measured attributes have to be matched to the model. The main contribution of this research is the methodology for the generation of rules for model matching or classification. The a priori probability distribution of attributes in each group is used to generate the rules for the model that corresponds to the group. Rules generated for all models are collected in a rule base and used for classification. A systematic method for integrating expert and heuristic knowledge is developed in order to improve the efficiency of the classification process. An extensive process of error analysis of the present methodology is also presented. The techniques developed in the present research were implemented in a prototype fish processing cell that is being developed in the Industrial Automation Laboratory. This cell consists of a vision station, a knowledge-based system, a robotic cutter, and a motorized conveyor unit. In a heading (head removal) operation, each fish on the conveyer is imaged at the vision station and the corresponding cutting contour is determined with the help of the knowledge-based system. This cutting contour is transformed into a drive signal for the robotic cutter. At the cutting station, a fish will be gripped and cut according to the trajectory generated for that particular fish. In the present prototype system, however, the robot draws the corresponding cutting contour on a board placed on the conveyor bed. The vision system, the cutter/gripper controls, and the conveyer controls communicate and synchronize the activities of various components in the fish processing system. Specific ways of transferring the new technology to the Canadian fish processing industry are currently being explored. This research also compares the performance of the developed rule-based methodology with other classification and estimation methodologies. These include neural networks, Bayesian inferences, k-nearest neighbour method, and multiple regression. The results of these comparisons are presented. Also, as a preliminary requirement for implementation, algorithms have been developed for on-line measurement of the attributes using low-level image analysis techniques.

Thesis/Dissertation

application/pdf

2008-09-17

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.

http://hdl.handle.net/2429/2154

Mechanical Engineering

University of British Columbia

UBCV

DSpace