An experimental comparison of the EM algorithm versus general optimization for combined image identification and restoration

Abstract

Summary form only given. The problem of learning the parameters needed for image restoration from the given noisy and blurred image has been addressed. The asymptotically optimal approach of finding the maximum-likelihood estimate of the parameters and then using this value of the parameter to construct the restoration filter has been taken. One way to do this is to iteratively solve the nonlinear problem of maximizing the a posteriori probability of the image given the blurred observations and also the unknown parameters. The ellipsoidal algorithm and the expectation-maximization (EM) algorithm have been used for this purpose. An experimental comparison of these two methods for parametrically restoring images when the parameters are not known a priori has been made.<>