An MRF-based image segmentation with unsupervised model parameter estimation

This paper deals with image segmentation when the image consists of uniform background (b.g.) and uniform foreground (f.g.) with noise. We formulate this problem into the joint minimization of MRF energy with respect to a label image and density parameters corresponding to f.g. and b.g., and solve it exactly in reasonable computation time. The proposed method efficiently solves the joint minimization by utilizing the novel property that multiple minimizations of MRF energy, corresponding to different combinations of density parameters for b.g. and f.g., can be solved by a single total-variation minimization. In addition, we also extend the proposed method to the case where label images together with density values corresponding to multiple smoothing (regularization) parameters can be obtained, exactly and simultaneously with a much shorter computation time compared with the trivial exhaustive search.