Identification of image structure by the mean shift procedure for hierarchical MRF-based image segmentation

Tree-structured Markov random fields have been recently proposed in order to model complex images and to allow for their fast and accurate segmentation. By modeling the image as a tree of regions and subregions, the original K-ary segmentation problem can be recast as a sequence of reduced-dimensionality steps, thus reducing computational complexity and allowing for higher spatial adaptivity. Up to now, only binary tree structures have been considered, which simplifies matters but also introduces an unnecessary constraint. Here we use a more flexible structure, where each node of the tree is allowed to have a different number of children, and also propose a simple technique to estimate such a structure based on the mean shift procedure. Experiments on synthetic images prove the structure estimation procedure to be quite effective, and the ensuing segmentation to be more accurate than in the binary case.