Modeling image textures by Gibbs random fields

Abstract Drawbacks of the traditional scenario of image modeling by Gibbs random fields with multiple pairwise pixel interactions are outlined, and a more reasonable alternative scenario based on Controllable Simulated Annealing is described. The latter scenario uses an analytic and stochastic approximation of Gibbs potentials to minimize a distance between the selected gray level co-occurrence or difference histograms for a given training sample and the simulated images.