Image Segmentation Based On Brushlet Domain HMT Model

Brushlet Transform is proposed by Meyer and Coifman in 1997. As a kind of Multiscale Geometric Analysis tools, it is sensitive to the orientation information of the image. Brushlet Transform has achieved wide application in image compression and image classification. Transform-domain statistical image modeling is based on a well-grounded theory, which has high percision and low complexity. It becomes a central issue in recent years.The statistical properties of Brushlet domain coefficients is analyzed in the paper. After that the Brushlet Domain Hidden Markov Model is constructed. Firstly, the concept of Multi-scale Brushlet Transform is proposed. The image is decomposed into each scale of the Brushlet domain, and the correlation between the Brushlet coefficients in adjacent scales is found:according to the orientation information, the energy in low scale sub-bands would be assigned to high scale sub-bands. And there is an interscale persistence of the energy distribution between the coefficients. In order to model the persistence, a "four to four" tree model with markov property is proposed in the paper. By analysising the distribution of Brushlet energy coefficients in the same scale, we found that it meets the Rayleigh distribution, and two states Rayleigh Mixture Model (RMM) can be used to make a good approximation to the distribution. On one hand, the RMM is used to finish the intra-scale modeling, on the other hand, the Hidden Markov Model is used to model the inter-scale modeling. Then, Brushlet Domain HMT (BruHMT) model is constructed. At last reference to the Wavlet Domain HMT (WHMT) image segmentation method, we apply the BruHMT method to texture image segmentation and obtain a good segmentation results. Especially for coarse texture segmentation, the segmentation accuracy of BruHMT method is significantly higher than the WHMT segmentation method.There are also some shortcomings in the BruHMT based method, such as edge positioning inaccuracy and serious loss of image detail information. These issues are particularly prominent in the SAR image segmentation, for which we improved the BruHMT.①Using a strategy called intersacle fusion to process the raw segmentation obtained from BruHMT to take a compromise between the regional coherence and detail information②the Watershed method is proposed, and the advantage of Watershed is used to improve the edge segmentation of our method. The experimental results show that the improved BruHMT method is effective.