Automatic texture segmentation based on wavelet-domain hidden Markov tree

Abstract

An automatic texture segmentation approach is presented in this paper, in which wavelet-domain hidden Markov tree (WD-HMT) model is exploited to characterize the texture features of an image, an effective cluster validity index, the ratio of the overlap degree to the separation one between different fuzzy clusters, is used to determine the true number of the textures within an image by solving the minimum of this index in terms of different number of clusters, and the possibilistic C-means (PCM) clustering is performed to extract the training sample data from different textures. In this way, unsupervised segmentation is changed into self-supervised one, and the well-known HMTseg algorithm in the WD-HMT framework is eventually used to produce the final segmentation results, consequently automatic segmentation process is completed. This new approach is applied to segment a variety of composite textured images into distinct homogeneous regions with satisfactory segmentation results demonstrated. Real-world images are also segmented to further justify our approach. © Springer-Verlag Berlin Heidelberg 2005.