Parallelization of Low-Level Computer Vision Algorithms on Clusters

Abstract

In this paper we present parallel implementations of some representative low level vision algorithms on a cluster of workstations. These include convolution operation and the image restoration algorithm using Markov random field models. The convolution operation has been parallelized using the Farmer-Worker paradigm, while the image restoration algorithm has been parallelized through the Master-Worker pattern. Parallel implementations of both these algorithms have shown promising results, where the observed speedups are reasonably close to the ideal speedups. The parallel convolution operation has shown good scalability with respect to the problem size and number of processors used in parallelization. The paper elaborates on different parallelization results of the convolution operation observed after varying the image size, mask size and processor load on individual workstations. The image restoration algorithm is communication intensive. However, as the computing time between successive communications at each worker process is relatively high, the actual speedups observed are very close to the ideal speedups in this algorithm.