Maximum likelihood image identification and restoration using genetic algorithms

A wide range of images can be modelled as noisy observations of a 2D auto regressive-moving average (ARMA) process. The AR part of the model describing the underlying ideal image statistics and the MA part the characteristics of the imaging system. In restoring the image it is desired to remove the effects of the imaging system. This task is often complicated by the fact that the coefficients which describe the characteristics of the ARMA model and occluding noise are unknown a priori and must be estimated from the degraded image. In recent years several methods for solving this image identification problem based on maximum likelihood estimation techniques have been proposed. These techniques are principally gradient based and often, because of the multimodal nature of the problem, fail to converge correctly. In this paper it is illustrated, with practical examples, the application of genetic algorithms to this problem