Image Segmentation Based on GBP Algorithm

Abstract

Belief propagation (BP) algorithm is an efficient way for image segmentation based on graphical models. However BP fails to converge when the graph has cycles. Generalized belief propagation (GBP) provides more accurate solutions on such graphs. In this paper, a method based on GBP algorithm is proposed for image segmentation. In proposed method, class label is modeled using Gaussian Markov random fields (GMRF), and expectation maximization (EM) algorithm was adopted to estimate the hyper-parameters of GMRF. After region graph constructed, we run GBP algorithm on region graph, to maximize the posteriori conditional probability distribution based on Bayesian theory. The analysis and experiments on natural images showed that it gives much more accurate results than those found using ordinary belief propagation.