Image Segmentation Based on Bethe Approximation for Gaussian Mixture Model

We propose an image segmentation algorithm under an expectation-maximum scheme using a Bethe approximation. In the stochastic image processing, the image data is usually modeled in terms of Markov random fields, which can be characterized by a Gibbs distribution. The Bethe approximation, which takes account of nearest-neighbor correlations, provides us with a better approximation to the Gibbs free energy than the commonly used mean-field approximation. We apply the Bethe approximation to the image segmentation problem and investigate its efficiency by numerical experiments. As a result, our approach shows better robustness and faster converging speed than those using the mean-field approximation.