Many inverse problems focus on recovering a quantity of interest that is a priori known to exhibit either discontinuous or smooth behavior. Within the Bayesian approach to inverse problems, such structural information can be encoded using Markov random field priors. We propose a class of priors that combine Markov random field structure with Student's t distribution. This approach offers flexibility in modeling diverse structural behaviors depending on available data. Flexibility is achieved by including the degrees of freedom parameter of Student's t distribution in the formulation of the Bayesian inverse problem. To facilitate posterior computations, we employ Gaussian scale mixture representation for the Student's t Markov random field prior, which allows expressing the prior as a conditionally Gaussian distribution depending on auxiliary hyperparameters. Adopting this representation, we can derive most of the posterior conditional distributions in a closed form and utilize the Gibbs sampler to explore the posterior. We illustrate the method with two numerical examples: signal deconvolution and image deblurring.
View on arXiv