Reconstruction of Gray-scale Images

Abstract

Reconstruction of images corrupted by noise is an important problem in Image Analysis. In the standard Bayesian approach the unknown original image is assumed to be a realization of a Markov random field on a finite two dimensional finite region. This image is degraded by some noise, which acts independently in each site of and has the same distribution on all sites. The reconstructed image is the maximum a posteriori (MAP) estimator. When the number of colors is bigger than two the standard a priori measure is the Potts model, but with this choice reconstruction algorithms need exponential growing time in the number of pixels. We propose a different a priori measure that allows to compute the MAP estimator in polinomial time. The algorithm binary decomposes the intensity of each color and reconstructs each component, reducing the problem to the computation of the two-color MAP estimator. The latter is done in polinomial time solving a min-cut max-flow problem in a binary graph as proposed by Greig, Porteous and Seheult (1989).