A generative perspective on MRFs in low-level vision

摘要: Markov random fields (MRFs) are popular and generic probabilistic models of prior knowledge in low-level vision. Yet their generative properties rarely examined, while application-specific non-probabilistic learning gaining increased attention. In this paper we revisit the aspects MRFs, analyze quality common image priors a fully application-neutral setting. Enabled by general class MRFs with flexible potentials an efficient Gibbs sampler, find that do not capture statistics natural images well. We show how to remedy exploiting sampler for better based on potentials. perform restoration these computing Bayesian minimum mean squared error estimate (MMSE) using sampling. This addresses number shortcomings have limited so far, leads substantially improved performance over maximum a-posteriori (MAP) estimation. demonstrate combining our learned sampling-based MMSE estimation yields excellent application results can compete recent discriminative methods.