Low level image partitioning guided by the gradient watershed hierarchy

In this paper, we study the construction of hierarchies for gradient waters heds. The underlying analysis is based on the fact that the gradient image provides information with respect to local discontinuity and similarity. Tw o algorithms which utilize this type of information are analyzed. The first one is an extended version of the dynamics of contours and the second is t he algorithm which considers the minimum cost path saliency between two adj acent regional minima. For both algorithms, a stopping criterion which defi nes and extracts automatically each hierarchical level is proposed. An impo rtant hint to determine which one of the two algorithms has a superior beha vior is the domination of the profile type of the surface (convex or concav e) formed by the gradient image. Tests are discussed on artificially genera ted and real images. The proposed bottom-up procedure provides an essential aid to the precise classification of anatomical objects when used together with a user-interface environment. (C) 1999 Elsevier Science B.V. All righ ts reserved.