Scheduling techniques for parallel implementation of wear particle recognition algorithms

Many schemes in the field of computer vision and image processing, present potential for parallel implementations through one of the three major paradigms: geometric parallelism, algorithmic parallelism, and processor farming. Static process scheduling techniques are used successfully to exploit geometric and algorithmic parallelism, whilst dynamic process scheduling is better suited in dealing with the independent processes inherent in the processor farming paradigm. This paper considers the application of parallel, or multicomputers to a class of problems exhibiting spatial data characteristic of the geometric paradigm and is best suited to a static scheduling scheme. However, by using the processor farming paradigm, a dynamic scheduling technique is developed to suit the MIMD structure of the multi-computers. The specific problem chosen for the investigation is the recognition and classification of microscopic wear particles generated by wear mechanisms. Experiments are performed on static and dynamic schemes, and are compared in terms of total processing time, speedup, and efficiency.