Multiscale statistical image models and Bayesian methods

Multiscale statistical signal and image models resulted in major advances in many signal processing disciplines. This paper focuses on Bayesian image denoising. We discuss two important problems in specifying priors for image wavelet coefficients. The first problem is the characterization of the marginal subband statistics. Different existing models include highly kurtotic heavy-tailed distributions, Gaussian scale mixture models and weighted sums of two different distributions. We discuss the choice of a particular prior and give some new insights in this problem. The second problem that we address is statistical modelling of inter- and intrascale dependencies between image wavelet coefficients. Here we discuss the use of Hidden Markov Tree models, which are efficient in capturing inter-scale dependencies, as well as the use of Markov Random Field models, which are more efficient when it comes to spatial (intrascale) correlations. Apart from these relatively complex models, we review within a new unifying framework a class of low-complexity locally adaptive methods, which encounter the coefficient dependencies via local spatial activity indicators.