Bayesian image modeling is presented based on a generalized sparse prior probability distribution. Our prior includes sparsity in each interaction term between every pair of neighbouring pixels in Markov random fields. A new scheme for hyperparameter estimation is based on the conditional maximization of entropy in our generalized sparse prior. In addition, the criterion used for defining the optimal value for sparseness in interactions is that of the maximization of marginal likelihood. Our practical algorithm is based on loopy belief propagation.
- Bayesian statistics
- Belief propagation
- Markov random fields
- Maximum likelihood estimation
- Probabilistic image processing
- Statistical-mechanical informatics