Recently, vision research has centred on both the extraction and organization of geometric features, and on geometric relations. It is largely assumed that topological structure, that is linked edgel chains and junctions, cannot be extracted reliably from image intensity data. In this paper we demonstrate that this view is overly pessimistic and that visual tasks, such as perceptual grouping, can be carried out much more efficiently and reliably if well-formed topological structures are available. The widespread assumption that edge detectors produce incomplete and erroneous topological relations, such as the image projection of polyhedral face-edge-vertex structures, is shown to be false by analyzing the causes for failure in traditional edge detectors. These deficiencies can largely be overcome, and we show that a good compromise between topological completeness and geometric accuracy can be achieved. Furthermore, edge detection should not be carried out in isolation. The resulti...

Abstract not found

Genetic Programming (GP) is a powerful machine learning technique derived from Genetic Algorithms. We use GP to produce high-performance edge detectors in 1dimensional signals. Using the theory of edge detection to assess candidate detectors produced by GP, we show that this technique can produce detectors that often outperform the established `optimal' detectors. Further, we show that detectors can be evolved to be specific to the training data, offering the capability to produce tailor-made edge detectors. Such detectors are shown to perform better on the data set than other evolved detectors. 1. Introduction Edge detection is a problem to which a large amount of effort has been devoted. It is an essential first step in many machine vision systems, for example, extracting important image features as the first stage in the signal-to-symbol process. Detecting edges well is therefore an - 2 - important problem. In addition to a considerable body of work on applying different practical...

The extended Euclidean distance transform Shape representation has always played a central role in computer vision. Skeletal shape descriptors which make symmetry explicit are an important class of shape repre- sentations. The goal of this thesis is to study the problems encountered using skeletal shape descriptors. The thesis unites three main themes of work: A filter based approach to skeletonisation, skeletonisation using parallel wave propagation and skeletonisation using an extended Euclidean distance transform. The distance transform approach to skeletonisation computes a skeleton by identi- fying the so called local maxima of the distance transform. A new method has been proposed to detect these features using a filter-base approach inspired by models of processes in the human visual system. Further improvements were made by using a filter designed to detect a specific geometric feature on the distance transform which corresponded to the skeletal points. This improved the quality of skeletons obtained but could only compute restricted skeletal descriptions. The wave propagation algorithm of Brady and Scott has been studied; they origi- nally implemented this on a simulator of the Connection Machine. The issues of map- ping the algorithm onto an array of transputers have been investigated. An efficient implementation was realised by reducing synchronisation and data transfer overheads. It was found that the algorithm could compute more general shape descriptions than the distance transform approach but the quality of skeletons produced was not as good. Using standard techniques from singularity theory, an analysis of distance functions from object boundaries has been undertaken. This resulted in the formal definition of a new extended Euclidean distance transform. An algorithm has been devised to perform skeletonisation using the extended distance transform. This combined the advantages of the filter and wave based techniques in that it produced skeletons of a high quality which made more symmetries explicit than the standard distance transform approach. In addition, the extended distance transform provides an elegant unifying framework for work on skeletal shape descriptions. Abstract Above all I would like to thank my supervisor Roberto Cipolla for his support, advice

Edge detection is the process of detecting discontinuities in signals and images. We apply genetic programming techniques to the production of highperformance edge detectors for 1-D signals and image profiles. The method, which it is intended to extend to the development of practical edge detectors for use in image processing and machine vision, uses theoretical performance measures as criteria for the experimental design.

Edge detection is the process of detecting discontinuities in signals and images. We apply Genetic Programming techniques to the production of high-performance edge detectors for 1D signals and image profiles. The method, which it is intended to extend to the development of practical edge detectors for use in image processing and machine vision, uses theoretical performance measures as criteria for the experimental design.