Data-Driven Segmentation of Textured Images Using Hierarchical Markov Random Fields.

This paper proposes an algorithm for a data-driven segmentation of an image consisting of multiple textures. The method uses hierarchical Markov random fields (MRF) consisting of two layers: the first layer MRF representing invisible regional images; and the second layer MRF representing a different texture in each region. This method is based on the EM algorithm for estimating the maximum likelihood of incomplete data, to treat invisible regional processes. The algorithm repeats an estimation of the MRF parameters and a segmentation (estimation of region processes). The former is carried out by maximizing a pseudolikelihood function; the latter is carried out by applying ‘the deterministic iterative relaxation algorithm’ (developed by the authors) to the textured images. The proposed algorithm has been confirmed successfully by computer simulations using artificially composed textures. [ABSTRACT FROM AUTHOR]