On combining higher-order MAP-MRF based classifiers for image labeling

Abstract

In this paper we present a framework for combining MAPMRF based classifiers for solving image labeling problems, by deriving a classification rule that uses a Gaussian Markov Random Field to model the observed data and a higher-order Potts MRF model as prior knowledge. In this scenario, the Pottsmodel parameter acts like a regularizarion parameter, controlling the tradeoff between data fidelity and smoothing. Maximum Pseudo-Likelihood equations are applied to automatically set this parameter value. The proposed methodology consists in using several initial conditions for the iterative combinatorial optimization algorithms in order to escape local maxima solutions. Experiments with NMR image data show, in quantitative terms, that the joint use of multiple initializations and higher-order neighborhood systems significantly improves the classification performance. © Springer-Verlag Berlin Heidelberg 2011.