Stochastic and deterministic algorithms for MAP texture segmentation

Tal Simchony, Rama Chellappa

Abstract: A model-based approach is proposed for the problem of texture segmentation using a maximum a posteriori (MAP) estimation technique. A Gauss-Markov random field (GMRF) is used for the conditional density of the intensity array, given the unobserved texture class and a second-order Ising distribution for the prior distribution over the texture classes. The GMRF model for the conditional density allows a closed-form expression for the density to be written, so that the dependence of the density on the label parameters can be expressed. This expression is used here to derive the joint distribution of intensity and label arrays. The joint distribution is maximized using the stochastic relaxation method and the deterministic iterated conditional mode (ICM) technique. The ICM algorithm can be implemented efficiently on a neural net with local connectivity and regular structure. Comparisons of these two methods are given using real textured images.<