Transputer configurations for computer vision

Computer analysis of static images imposes a significant computational burden on the processing hardware. In dynamic vision, the problem is manifold, and the requirement is also to reduce the latency of the processing, in order to allow realistic reaction times to events in the scene. Flexible, massively parallel architectures hold the promise of fulfilling these requirements for low, medium and high level vision tasks, provided that robust algorithms can be implemented in an efficient manner. In this thesis the role of the transputer as an intelligent processing element for multi-processor parallel architectures will be examined to determine its suitability across the spectrum of vision processing levels. To explore such possibilities, two fields within computer vision will be investigated. Initially, low and medium level vision tasks will be explored to apply the transputer to the field of label inspection. This investigation will include the introduction and analysis of a data-routing mechanism which will then be compared with one already popular in the world of transputers. Both techniques are suitable for geometric parallelism. In the course of these inquiries, a highly parallelised approach to the solution of the normal parameterisation of the Hough transform will be presented. Next, to investigate the more demanding aspects of computer vision, bordering around medium to high level vision, the ideas of temporal continuity and motion correspondence of image features in time-varying sequence of images will be examined. A parallel model is described which is designed for use as a basis for implementation of image-feature tracking algorithms on general parallel architectures. The model is in-dependent of feature tracking algorithms. An implementation of the model is outlined using a tracking algorithm founded on features such as the mid-point, orientation and the length of edge segments, and using a modified form of the Kalman filter. The implementation consists of three independent units each of which has been applied in a studied transputer configuration. For example, the tracking unit is based on a tree configuration and displays MIMD characteristics. The edge extraction unit borrows from earlier work in the thesis and further investigates that approach. Overall, this thesis spans the fields of image processing and parallel processing in the investigation for the applicability of the transputer