Image recovery by simulated annealing
Abstract
The basic problem of image recovery and pattern recognition is to determine the original pattern, f, given its corrupted version, g. The unknown pattern, f, is an element of the set S = f1, f2, f3, and the task is to deduce which pattern in S gave rise to the image data, g. S is called the solution candidate space and could be, for example, the set of alphabetical symbols. If it is known that certain elements of S have a higher probability of occurring than others (such as alphabetical symbols in text), this a priori information can be incorporated into the procedure for finding f according to the techniques of Bayesian analysis. In the general image recovery problem, S is the set of all possible patterns on an n x n pixel image, and the relationship between the original image, f, and the image data, g, can be modeled by g = f + 2, where w is random noise. The set of pixel brightnesses at lattice position (i,j) are described by f = fij, g = gij, and w = wij. Image recovery and pattern recognition problems are thus combinatorial optimization problems, in which a solution candidate space, S, must be searched. The larger S is, the more difficult the search.
 Publication:

Final Report
 Pub Date:
 November 1990
 Bibcode:
 1990hdl..rept.....C
 Keywords:

 Bayes Theorem;
 Combinatorial Analysis;
 Image Reconstruction;
 Pattern Recognition;
 Brightness;
 Pixels;
 Random Noise;
 Optics