An Evaluation of the ICM (Iterated Conditional Modes) Algorithm for Image Reconstruction

In the last few years considerable interest has been shown in the problems posed by the analysis of images corrupted by random noise. The reconstruction of such images leads to special difficulties as it is an ill- posed problem in the sense described by OSullivan, 1986. Typically the reconstruction of an array of pixels will have as many parameters as observations. A number of techniques have been proposed which solve ill-posed problems by restricting the class of admissible solutions, see Marroquin, Mitter Poggio 1987. This is achieved by introducing a priori knowledge about admissible solutions. Much interest currently centres on techniques which incorporate knowledge about the underlying image using Bayesian methodology, See Geman Geman 1984 Kashyap Lapsa 1984. These techniques assume that the underlying scene can be adequately described as a realisation from a prescribed Markov random field. Motivated by this approach Besag 1986 introduced a technique known as iterated conditional modes ICM. This iterative procedure incorporates knowledge about the underlying scene by the choice of a neighborhood system, weight function and smoothing parameter. Broadly speaking this method exploits the tendency of adjacent pixels to have the same colour. A similar approach based on spatial auto regression is described in Woods, Dravida Mediavilla 1987. RH