On Optimization of Probability Vector Random Fields Used for Image Segmentation

In the context of image segmentation, Markov random fields (MRF) are extensively used. However solution of MRF-based models is heavily dependent on how succesfully the MRF energy minimization is performed. In this light two methodologies, complementary to each other, are proposed for optimization of the special class of models comprising of a random field imposed on label priors. This class of segmentation models poses a special optimization problem, as the variables constituting the MRF in this case are continuous and are subject to probability constraints (positivity, sum-to-unity). The proposed methods are evaluated numerically in terms of objective function value and segmentation performance, and compare favorably to existing corresponding optimization schemes.