Multiscale watershed segmentation of multivalued images

In this paper, a new segmentation technique for multivalued images is elaborated. The technique accesses multiscale edge information of a multivalued image by a concept, called multiscale fundamental form. At different scales, an edge map of the multivalued images is obtained, at which a watershed-based algorithm is applied. The multiscale behavior of the obtained watershed regions is employed to conduct a region merging procedure. In order to remove noise or local texture, prior to segmentation an anisotropic diffusion filter is applied, also making use of the multiscale fundamental forms. In this way, the entire procedure is applied using multivalued processing.