City Research Online

Abstract

The work in this thesis is concerned with parallel architectures based on the Inmos transputer-type processors and parallelisation of some computer vision tasks chosen from low to high level.

The transputer is a microprocessor with a micro-programmed scheduler and four serial communication links. It directly supports parallel processing since several transputers can be connected through their links to co-operate on solving a problem. Also several processes can be run on the same transputer. A major issue in parallel processing is the communication overhead introduced by parallelising a given task. This overhead is not present in sequential processing and must be curbed if the implementation of a task on a parallel machine is to be successful. The interconnection network underlying the architecture of a parallel computer is therefore of the utmost importance.

Computer Vision consists of a hierarchy of tasks ranging from low-level operations dealing with large amounts of relatively simple data to high level operations handling increasingly complex structures. In this work a novel edge detector based on adaptive filtering and an edge detector operating on colour images are presented and implemented on a number of transputers. These parallel implementations together with implementations of vector quantisation, Fourier descriptors for shape discrimination, the Hough transform and the Maximum clique algorithm, offer a notable performance increase when compared with sequential implementations. However, every algorithm required the design of a specific network of transputers to take advantage of the parallelism and data dependencies inherent in each.

Consequently, attention is focused on the topology of interconnection networks. In particular, the communication requirements of computer vision algorithms as identified by the various computer vision tasks are analysed. These requirements together with graph theoretical considerations are then used to suggest a topology for large transputer networks. The latter is based on sub-graphs, with proven performance when used to implement interconnection networks, combined to form an architecture with improved performance. This architecture consists of a fixed structure supplemented with a dynamically reconfigured network. After describing this topology, a routing algorithm that conveys messages along shortest paths in the network is given and implemented. And finally, some practical issues in the use of transputers are considered and solutions proposed.

Publication Type:	Thesis (Doctoral)
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Departments:	School of Science & Technology > Computer Science School of Science & Technology > School of Science & Technology Doctoral Theses Doctoral Theses

Export

Downloads