Morphological segmentation revisited

Abstract

Morphological segmentation is now 25 years old, and is presented in textbooks and software libraries. It relies first on the watershed transform to create contours and second on markers to select the contours of interest [12]. Separating the segmentation process into two parts constitutes its main interest: finding the markers is the intelligent part, since it requires an interpretation of the image, but the difficulty is counterbalanced by the fact that the resulting segmentation is largely independent of the precise shape or position of the markers. Furthermore, the tedious part, that is finding the contours is entirely automatic. The advantages of the method are good localization, invariance to lighting conditions, absence of parameters, high sensitivity as strongly or weakly contrasted objects are equally segmented. It has been used with success in many circumstances, in any number of dimensions and has become extremely popular. Searching the webpages containing the word watershed and segmentation, Google finds more than 15000 pages ! Another reason for its popularity is the speed of the watershed transform : hierarchical queues allow to mimic the flooding of a topographic surface from a set of markers, and require only one pass through the image [3, 16, 22]. The watershed has also successfully been implemented on dedicated hardware, on DSPs and on parallel architectures [8, 4]. Morphological segmentation being a success story, is it necessary to devote a new paper to a method so widely known ? Indeed yes, for the research in the domain is more active than ever.