AN ADAPTIVE FINITE-ELEMENT PROCEDURE FOR THE IMAGE SEGMENTATION PROBLEM

The image segmentation problem in computer vision is considered. Given a two-dimensional domain D and a function defined on it (the original image), the problem is to obtain a 'cartoon' associated with this fun ction, namely to find a set of inner boundaries which divide D into su bdomains (objects) in an optimal way. The optimality criterion used he re is given by the Mumford-Shah (MS) and Blake-Zisserman model, which leads to a strongly non-linear problem. Related problems appear in mul tiphase continuum mechanics. An iterative procedure based on an h-adap tive finite element method is proposed for the solution of this proble m. The mesh adaptivity enables an efficient solution technique, with t he use of basic coarse discretization and a few local regions of high resolution where needed. (C) 1998 John Wiley & Sons, Ltd.