Multiresolution Watershed Segmentation on a Beowulf Network

Among the many existing multiresolution algorithms, the scale-space approach offers the benefits of strong mathematical and biological foundation and excellent results, but the serious drawback of a heavy computational load. Parallel implementation of this category of algorithms has never been attempted. This article presents a first experiment, using the multiresolution watershed segmentation as algorithm, and an 8-node Beowulf network as hardware platform. First, the classical approach is followed whereby the image is divided in several regions that are separately allocated to different nodes. Each node performs all the calculations for his region, at any level of resolution. Next, a truly multiresolution approach is followed, allocating the workload to the processors per resolution levels. Each node is allocated a number of resolution levels in the scale space, and performs the calculations over the whole image for the particular resolution levels assigned to it. The implementation in the latter approach is clearly much more straightforward, and its performance is also clearly superior. Although the experiments using the region-wise assignment were only done by splitting up the image in rows, and not in columns or in quadrants, the difference in the results is so dramatic that the conclusions can easily be generalized, pointing to the fact that scale space algorithms should be paralellised per resolution level and not per image region.