Texture image segmentation by hierarchical modeling

The Texture Fragmentation and Reconstruction (TFR) algorithm has been recently proposed for the unsupervised hierarchical segmentation of textures. It is based on a hierarchical image model, where textures are characterized in terms of their spatial interaction properties, modeled by means of a set of Markov chains, each one associated with a major spatial direction. The TFR algorithm fits the image to the hierarchical model by means of a split-and-merge procedure where the first step (fragmentation) aims at extracting the elementary texture states, which are progressively merged in the second step (reconstruction), so as to obtain a hierarchical nested segmentation. Although TFR results are usually very good, it has been sometimes observed a bias towards the undersegmentation for complex images. Here, we analyze this phenomenon and propose the use of an improved fragmentation step, where would-be elementary states are ranked based on a suitable measure of their reliability and possibly purged. Experimental results validate the effectiveness of the new algorithm.