Regional category parsing in undirected graphical models

This article presents an approach for regional categorization in complex natural scenes with undirected graphs. A novel MRF-like model is proposed with spatial constraints in the feature space based on existing directed graphs, and an approximation of pseudo-likelihood is introduced for probability inference and parameter estimation. With this approximation, we can deal with the intractability of potential functions and get spatial relations between patches of different classes for more information in their co-occurrence matrix. The Receiver-Operating-Characteristic curves in our experiments demonstrate a better performance from our proposed method in comparison with directed probabilistic models such as LDA and constellation.