Unsupervised multiscale image segmentation using wavelet domain hidden Markov tree

In this paper, we have improved the supervised multi-scale texture segmentation (HMTseg), where wavelet domain hidden Markov model is applied to capture the texture feature and a contextual model is employed to fuse multi-scale segmentation. In order to extend supervised HMTseg to an unsupervised one, we perform a hierarchical clustering in the blocks on the starting scale. The dissimilarity between sub-mages is measured by the Kullback-Leibler distance (KLD) between corresponding WD HMT models. Experiments show that the performance of our proposed method is promising and needs less prior information on textures present in the given images.