Load Redistribution Algorithms for Parallel Implementations of Intermediate Level Vision Tasks

Parallelism can be exploited to handle the enormous computational requirements of many vision applications. However, the computational power oered by multiprocessor architectures cannot be fully harnessed to achieve the desired speedup. This is primarily due to the unbalanced distribution of computational load among the processors of a parallel architecture. Furthermore, in parallel implementations of image analysis tasks, what constitutes computational load and the load balancing requirements of specic implementations are often dicult to dene in a systematic way. In this paper, we consider the load balancing requirements of parallel implementations of intermediate level vision tasks on distributed memory parallel architectures. The computational characteristics of such tasks are brie BLOCKINy discussed and an appropriate denition of computational load is adopted. The primary implication of this denition for load balancing is that load entities to be redistributed are allowed to have nonuniform computational cost. An existing algorithm, which assumes uniform cost loads, and two modications of this algorithm, which handle the nonuniform cost loads encountered in parallel implementations of intermediate level vision tasks, are described. These algorithms have been implemented on the iPSC/2 hyper-cube and their performance has been evaluated using simulated load conditions, as well as in the context of a simple object recognition system. Results on load balancing accuracy and total execution time are presented and discussed. Algorithmic performance has also been compared with the cases of optimal load distribution and no load redistribution. This work emphasizes the importance of understanding the requirements and diculties of load redistribution in parallel image processing applications. In this paper, parallel image processing is approached from the viewpoint of data parallelism. To achieve signicant speedups by exploiting data paral-lelism, one must properly balance the computational load of available processors. This is particularly true in the case of intermediate level vision tasks, which are characterized by a strong dependence of load distribution on image content. At the early stages of computation, image data is uniformly distributed among all available processors. However, as the computation on image data proceeds, specic image features emerge and these may be nonuniformly distributed on the image plane. Consequently, for further computations on these features, the computational load distribution may also be nonuniform. In order to maintain a high eciency of task execution, image features and the corresponding computational load must be properly balanced through redistribution to all available processors. Load redistribution algorithms are characterized as or. Static redistribution algorithms assume …