EM algorithm for bounded generalized t mixture model with an application to image segmentation

This paper develops an EM-type algorithm for fitting the bounded generalized t mixture (BGTM) model. Due to heavy tails, high kurtosis and bounded nature, the BGTM model provides a flexible and suitable model for many computer vision and pattern recognition problems. We develop a feasible expectation conditional maximization (ECME) algorithm for computing the maximum likelihood estimates of model parameters via selection mechanism. To validate the effectiveness of the proposed methodology, we conduct experiments on both simulated data and real natural images. The obtained results demonstrate that the model and the estimation algorithm outperforms its sub-models in terms of both accuracy and computational efficiency.