MRF Model-Based Restoration of Multiple-Texture Images

Abstract

In most MRF-based Bayesian restoration algorithms, the image is modeled by a single MRF. However, an MRF is a proper model only for simple images such as piecewise constant or homogeneous ones having the same statistics over the entire image. This hamper the applicability of these algorithms to more complex images. To overcome this shortcoming, we employ a hierarchical triply stochastic process to model the observed image and develop an iterative algorithm for the restoration of images with region-dependent statistics. The algorithm is developed for two cases of binary and gray level images degraded by flip noise and Gaussian white noise, respectively. No prior knowledge of the noise parameters or the parameters of two hidden processes that model the true image is assumed. The algorithm is data-driven except for the number of regions which is assumed known. The proposed algorithm also provides a segmentation of the observed noisy image as a byproduct. Some simulation examples showing the effectiveness of the algorithm are presented.