Image artifact reduction using maximum likelihood parameter estimation

Abstract

A technique for post-processing decoded compressed images to reduce decoding-related artifacts employs a maximum likelihood estimation of an original image f. The decoded image is modeled as a montage of “flat surfaces” of different intensities, where the number of flat surfaces and their intensities are generally different in different regions of the decoded image. The intensity of each pixel is conditionally adjusted to that of a corresponding flat surface in a window region surrounding the pixel. In a general algorithm, the flat surface model is fitted to the observed image by estimating the model parameters using the “k-means” algorithm and a hierarchical clustering algorithm. A cluster similarity measure (CSM) is used to determine the number of intensity clusters, and hence flat surfaces, in the model of a window region surrounding a pixel of interest. The pixel intensity is adjusted to an estimated value which is the mean intensity of the cluster in which the pixel falls. A simplified version of the method employs a three-cluster model in which the cluster centers are initialized by a deterministic rule. This simplified method is non-iterative in nature, thus requiring fewer computational resources.