In this paper, the problem of edge preserving smoothing in image processing is tackled by combining a noise corruption model and a region and edge image model. The derivation of the probability model for the first order difference in the gray levels of the region pixels and edge pixels lead to a non-linear filter with coefficients as functions of the estimated noise variance and edge intensity. Such a model-based approach allows the design of improved filters for noise filtering and image compression. Experimental results demonstrate the improved performance of the filter for both synthetic and natural images.