Bayesian image modeling and phase transition in generalized sparse Markov random fields and loopy belief propagation

by Kazuyuki Tanaka

Monday 16 Sept 2013, 14:30 → 16:30 Europe/Rome

Aula Careri (Dipartimento di Fisica, Ed. G. Marconi)

Bayesian image modeling is given for based on a generalized sparse prior probability. Our prior includes a sparsity in each interaction term between every pair of neighbouring pixels in Markov random fields. A new scheme for hyperparameter estimations is given from the standpoint of the conditional maximization for entropy in our generalized sparse prior. The criteria of the optimal value for sparseness in interactions is also given by using the maximization of marginal likelihood. Our practical algorithm is constructed by using the loopy belief propagation. Moreover, we discuss the first order phase transition in our generalized sparse Markov random fields by the loopy belief propagation.