Shape and Topology Constrained Image Segmentation with Stochastic Models

Abstract

The central theme of this thesis has been to develop robust algorithms for the task of image segmentation.<br /> All segmentation techniques that have been proposed in this thesis are based on the sound modeling of the image formation process. This approach to image partition enables the derivation of objective functions, which make all modeling assumptions explicit.<br /> Based on the Parametric Distributional Clustering (PDC) technique, improved variants have been derived, which explicitly incorporate topological assumptions in the corresponding cost functions.<br /> In this thesis, the questions of robustness and generalizability of segmentation solutions have been addressed in an empirical manner, giving comprehensive example sets for both problems. It has been shown, that the PDC framework is indeed capable of producing highly robust image partitions.<br /> In the context of PDC-based segmentation, a probabilistic representation of shape has been constructed. Furthermore, likelihood maps for given objects of interest were derived from the PDC cost function. Interpreting the shape information as a prior for the segmentation task, it has been combined with the likelihoods in a Bayesian setting. The resulting posterior probability for the occurrence of an object of a specified semantic category has been demonstrated to achieve excellent segmentation quality on very hard testbeds of images from the Corel gallery.