Distributed propagation of a-priori constraints in a Bayesian network of Markov random fields