Table I Psnr (db) Results for Image Corrupted by Gaussian and Impulse Noise on " the Convergence of Mean Field Procedures for Mrf's "

A new image approximation scheme is proposed. The structural constraints are incorporated in an iterative M -estimator algorithm. As a result, an image modeling method is obtained that is not influenced by outliers and reduces Gaussian and heavy-tailed noise efficiently; and, at the same time it retains important details. The images are modeled as tensor product bicubic B-splines. The smoothing parameter is estimated separately for each processing window, thus allowing it to adapt to local structures of the image. As a result, one can expect excellent Gaussian noise removal in smooth and slowly varying areas where is large and at the same time very good preservation of important details (small values of ). Results obtained by applying the filter based on the presented approximation algorithm to real scenes indicate that this method is robust with respect to variations in the statistics of both the noise and the image.