We discuss the problem of modeling structured textures and image data with random field models. Random field models have recently been applied to a variety of problems such as image enhancement, segmentation, and coding. The models have been shown to successfully reproduce stochastic patterns, but have not been as successful for capturing structure in patterns. Two possible solutions to the structured pattern problem are examined in this paper. Both solutions are based on a relationship between ``internal'' random field interactions and coupling to a set of ``external'' constraints. This relationship is also used to explain the performance of the Gibbs random field for some ``structured'' image segmentations.
PDF . Full list of tech reports