Approximate inference methods for grid-structured MRFs

Abstract

In this thesis, I compared the mean field, belief propagation, and graph cuts methods for performing approximate inference on an MRF. I developed a method by which the memory requirements for belief propagation could be significantly reduced. I also developed a modification of the graph cuts algorithm that allows it to work on MRFs with very general potential functions. These changes make it possible to use any of the three algorithms on medical imaging problems. The three algorithms were then tested on simulated problems so that their accuracy and efficiency could be compared.