Enhancements to the JPEG implementation of block smoothing method

Abstract

Summary form only given. This paper proposes several enhancements to the AC prediction approach, adapted by the Joint Photographic Expert Group (JPEG), for reduction of the blocking artifact effects. Our decoder uses value of reconstructed pixels of the already decoded part of the image, instead of the their DCT components. The major contribution of the paper is that we divide the prediction of DCT coefficients in two parts. For the low frequency coefficients, we solve a minimization problem. Its objective is to reduce the block boundary edge variance (BEV). The problem is solved analytically and its solution predicts DCT coefficients of a block in the terms of the first four coefficients of the four adjacent blocks. In this process, we also determine an optimal solution to the minimization of the mean squared difference of slopes (MSDS) considered for the same problem and computed using a quadratic programming method. For the mid-range frequency coefficients, we follow the interpolation method and interpolate image segments by ternary polynomials (JPEG uses quadratic polynomials). The smallest possible 9/spl times/9 pixel image segments are considered for the prediction of coefficients of 8/spl times/8 blocks (JPEG considers 24/spl times/24 pixel segments). The paper presents a complete formulation of the prediction equations, not provided by the JPEG. The paper also proposes three new statistical criterion to measure block boundary discontinuities. All enhancements have been added to a JPEG software. Results of several experiments using this software are given to compare the performance of different implementations of the AC prediction approach.