Dynamic texture segmentation using spectral clustering based on CHMMs

Abstract

In this paper, we introduce the spectral clustering method based on Continuous Hidden Markov Model (CHMM) into dynamic texture (DT) segmentation. In order to characterize the DT, CHMMs are used to model all spatial subblocks of the DT. The initial segmentation is realized by utilizing the spectral clustering based on CHMMs. The similarity between two different CHMMs is measured with approximated Kullback-Leibler divergence (KLD). To improve the DT segmentation performance, the mathematical morphology method is also applied into further processing which is operated on the pixel level. Experimental results on artificially synthesized DT samples of DynTex dataset demonstrate the effectiveness of the proposed method.