Wavelet domain possibilistic c-means clustering based on Markov random field for image segmentation

In this paper, an unsupervised multiresolution image segmentation technique is presented, which combines wavelet domain Markov random field with possibilistic c-means clustering algorithm. At the determination of wavelet coefficients likelihood model stage, Gaussian mixture model is used to characterize the wavelet coefficients statistical distribution, and the model parameters are estimated by expectation maximization algorithm. In order to capture the clustering property of wavelet coefficients, we establish the prior rule, the optimum conditional probability likelihood model of wavelet coefficients given the labels is determined. At the image segmentation stage, we establish possibilistic c-means clustering objective function based on the conditional probability likelihood model of wavelet coefficients. In order to capture the clustering property of wavelet coefficients, we incorporate the local statistical distribution of wavelet coefficients into the clustering objective function. The improved objective function with spatial constraints is optimizated, we can get a new image segmentation algorithm. The simulation on magnetic resonance image shows that the new multiresolution image segmentation technique obtains much better segmentation results, such as the accuracy of boundary localization and the correctness of distinguishing different tissues.