MRF-based Algorithms for Segmentation of SAR Images

In this paper we demonstrate a new method for Bayesian image segmentation, with specific application to Synthetic Aperture Radar (SAR) imagery, and we compare its performance to conventional Bayesian segmentation methods. Segmentation can be an important feature extraction technique in recognition problems, especially when we can incorporate prior information to improve the segmentation. Markov Random Field (MRF) approaches are widely studied for segmentation, but they can be computationally expensive and, hence, are not widely used in practice. This computational burden is similar to that seen in the statistical mechanics simulation of simple MRF models such as certain magnetic models. Recently, Swendsen and Wang and others have had great success accelerating these simulations using so-called “cluster” Monte Carlo methods. We show that these cluster algorithms can provide speed improvements over conventional MRF methods when the MRF prior model has sufficient weight relative to the observation model.