Redesigning of JPEG statistical model in the lossy mode fitting distribution of DCT coefficients

Abstract

The JPEG statistical models in the lossy mode specify the procedures for converting the discrete cosine transform (DCT) coefficients into binary strings and context modeling in the case where the binary arithmetic coder called the QM-coder is employed as an entropy coder. The JPEG lossy mode establishes two statistical models, one for prediction residuals of the DC coefficients and the other for the AC coefficients. We redesign these two models by taking account of their distribution. We confirm that the Laplacian distribution is appropriate for both the DC coefficients and the AC coefficients through the Kolmogorov-Smirnov (KS) test; consequently, we propose statistical models that fit the Laplacian distribution. By adopting the proposed statistical models in lieu of the conventional models, the number of the states decreases from 294 to 210 and the compression performance on several test images including super high definition images improves by 0.01 to 1.48%.