Image segmentation by label anisotropic diffusion

Weighing the difficulties of a symbolic description of 3D surface based scattered data, this article propounds a formalisation of the segmentation in discrete labelling terms. Global consistency of the result is expressed as a constraint satisfaction problem. To solve this problem, the method we present is based on an anisotropic diffusion principle along two structures respectively denoted minimal and maximal escarpment trees. These structures are drawn from the graph theory. Novel aspect of our method is its ability to work on non organised points and to detect arbitrary topological types of features, as crease edge or boundaries between two smooth regions. The proposed approach makes possible an hybrid segmentation, involving the duality between regions and boundaries. The method has proven to be effective, as demonstrated below on both synthetical and real data.