Restoration of Special Features in Lattice Images

The problem of interest is restoring special features of a degraded spatial pattern. An underlying unknown pattern is defined on a lattice. We modify the simulated annealing algorithm to produces under a practical annealing schedule a convergent sequence of probability distributions on the images. To this end, jumping probabilities of the annealing algorithm are incorporated without randomization. The proposed technique can also be viewed as an extension of the iterated conditional modes idea.

By a feature we mean a certain trait defining an image up to a smaller number of parameters. Two cases are distinguished. First, we assume that the presence of features is uncertain and is given some prior probability. We show how jumping probabilities can be defined to obtain a good restoration and also to choose the most likely feature. In the second case, the presence of features is assured and the restoration of the parameters specifying the feature is carried out.

Our restoration technique is empirically shown to produce fewer classification errors than a standard technique. In the second case, the restoration is necessarily in terms of the special features.