Adaptive Segmentation of Vector Valued Images

Abstract

In the recent years, various forms of integration of different visual cues within variational formulations based on the propagation of planae curves was considered for image segmentation as a step forward to early gradientbased methods. To this end, appearence-driven information as well as prior knowledge on the form of the structures of interest to be recovered were introduced to improve segemtnation performance. The main concern for using such modules is the requirement of having them formed off-line, either through pre-processing or defined by the user. In the recent literature, one can find segmentation methods that do not require neither intensity or shapebased prior knowledge. Going in this direction, in this chapter — driven by [413, 97] -, we propose a vector-valued segmentation method where the optimal partition and the statistical parameters of each class are dynamically recovered during the segmentation process.