Parallel multiresolution image segmentation with watershed transformation

A parallel extension for the watershed segmentation is presented in this paper. By following regional minima, i.e. the seeds around which regions are grown, in images of lower resolution, a region merging criterion in the oversegmented watershed image is retrieved. Thus, regions which have a common ascendent at a certain resolution level unify, as well as neighboring regions which annihilate until that level. Fine until coarse segmentation results are obtained efficiently in parallel on a Gray T3E parallel computer. Additionally, the technique performs quasi-automatically and allows an easy incorporation of restrictive merging criteria within the pyramidal merging structure.