Improving Potts MRF Model Parameter Estimation in Image Analysis

This paper presents a novel pseudo-likelihood equation for the estimation of the Potts MRF model parameter on second-order neighborhood systems. Experiments with simulated images comparing the proposed estimation method with a recent maximum likelihood estimation approach derived in literature show the superiority of our methodology. In order to evaluate the performance of the estimation method, we proposed a hypothesis testing approach to validate the obtained results. The test statistic together with the p-values, calculated through our approximation for the asymptotic variance of maximum pseudo-likelihood estimators, provide a complete framework for quantitative analysis of Potts model parameter estimation in image processing, pattern recognition and computer vision applications using MRF models.