Kadam, Sanjay S.; (1998) Design Patterns for Parallel Vision Applications. Doctoral thesis (Ph.D), UCL (University College London).

Text Design_patterns_for_parallel_v.pdf Download (8MB)

Abstract

Computer vision is a challenging application for high performance computing. To meet its computational demands, a number of SIMD and MIMD based parallel machines have been proposed and developed. However, due to high costs and long term design times these machines have not been widely used. Recently, network based environments, such as a cluster of workstations, have provided effective and economical platforms for high performance computing. But developing parallel applications on such machines involves complex decisions about distribution of processes over the processors, scheduling of processor time between competing processes, communication patterns, etc. Writing explicit code to control these decisions increases program complexity and reduces program reliability and code re-usability. We propose a design methodology based on design patterns which is intended to support parallelization of vision applications on a cluster of workstations. We identify common algorithmic forms occurring repeatedly in parallel vision algorithms and formulate these as design patterns. We specify various aspects of parallel behaviour of a design pattern, such as process placement or communication patterns, in its definition or separately as issues to be addressed explicitly during its implementation. Design patterns ensure program reliability and code re-usability since they capture the essence of working designs in a form that makes them usable in different situations and in future work. The research work is concerned with presenting a catalogue of design patterns to implement various forms of parallelism in vision applications on a cluster of workstations. Using relevant design patterns, we implement representative vision algorithms in low, intermediate and high level vision tasks. Majority of these implementations show promising results. For example, given a 512x512 image, the image restoration algorithm based on Markov random field model can be completed in less than 45 seconds on a network of 16 workstations (Sun SPARCstation 5). The same task takes more than 10 minutes on a single such workstation.

Download activity - last month

Export as CSVXMLJSON

Download activity - last 12 months

Export as XMLCSVJSON

Downloads by country - last 12 months

Export as JSONCSVXML

Archive Staff Only

View Item