Parallelising image processing algorithms

Multiprocessor machines provide increased computational power and memory capacity that can be used to achieve tasks involving large amounts of data, such as imaging. With multiprocessor machines, many sophisticated operations on image data can be accomplished within reasonable time constraints. In order to efficiently utilize multiprocessor machines, conventional image processing algorithms have to be parallelised. The design of parallel algorithms takes into account many considerations, e.g. interprocessor communication, load balancing, task division, task placement, scalability, network topology, etc. In this paper, the performance of some image processing algorithms running on a loosely-coupled multiprocessor machine is evaluated. The machine consists of a PC host computer and a multiprocessor network consisting of a number of transputers. The configuration of this transputer network is under software control and so different parallelisations of a particular algorithm can be tested on a particular network topology. Three image processing algorithms were selected for parallelisation. They are the Sobel edge operation, the fast Fourier transform and the Hough transform. Parallelism is achieved in various approaches, such as partitioning of tasks or partitioning of data. For a particular network configuration, the performance of different parallelisation approaches for each algorithm was assessed, based on the parallel processing time, overhead time, communication-to-computation ratio, efficiency, etc.<>