JPEG artifact removal using error distributions of linear coefficient estimates

Abstract

In this paper we present a method for JPEG artifact removal. The method works by estimating the distribution of a DCT coefficient given the values of the other coefficients, and then computing the expected value of this distribution in the quantization interval. We use information from an area exceeding the original block boundaries. Our method requires only information about image covariance, from which we estimate the effects of the transformations and quantization used in JPEG, under certain assumptions about the distributions. We will show that our method significantly improves the mean square error in our testing. Additionally, our method is shown to visibly reduce blocking artifacts in the images. © 2008 Springer-Verlag Berlin Heidelberg.