Parallel Split-Level Relaxation

Speed is a primary concern in all computer vision applications. With the advent of parallel processors it is necessary to explore the possibility of improving the performance of vision algorithms on these machines. Three avenues have to be pursued in this new framework. First, it is necessary to study the suitability of the existing algorithms on a multiprocessor. Secondly, new techniques to solve vision problems which are more amenable to parallel processing must be developed. Finally the algorithms must be implemented on a physical parallel processor to measure their true performance. This dissertation attempts to study the scene labeling problem focusing on these three aspects. A new technique, called split-level relaxation to solve this problem which is very suitable for parallel processing is given. Since the problem is NP-complete, several algorithms are used to reduce the average execution time. The arc consistency algorithms are analyzed for their performance on an abstract parallel machine. This gives an idea of the performance of the algorithms under ideal conditions. Finally, all the parallel algorithms are implemented on a BBN Butterfly Parallel Processor and results are also presented.