Fitting Markov Random Field Models to Images

摘要

We are interested in fitting two-dimensional, Gaussian conditional Markov random field (CMRF) models to images. The given finite image is assumed to be represented on a finite lattice of specific structure, obeying a CMRF model driven by correlated noise. The stochastic model is characterized by a set of unknown parameters. We describe two sets of experimental results. First, by assigning values to parameters in the stationary range, two-dimensional patterns are generated. It appears that quite a variety of patterns can be generated. Next, we consider the problem of estimating the unknown parameters of a given model for an image, and suggest a consistent estimation scheme. We also implement a decision rule to choose an appropriate CMRF model from a class of such competing models. The usefulness of the estimation scheme and the decision rule to choose an appropriate model is illustrated by application to synthetic patterns. Unilateral approximations to CMRF models are also discussed. (Author)