Optimal reconstruction value for DCT dequantization using laplacian pdf model

Abstract

Many image compression standards such as JPEG, MPEG or H.263 are based on the discrete cosine transform (DCT), quantization, and Huffman coding. Quantization error is the major source of image quality degradation. The current dequantization method assumes the uniform distribution of DCT coefficients. Therefore the reconstruction value is the center of each quantization interval. However DCT coefficients are regarded to follow Laplacian probability density function (pdf). We derive an optimal reconstruction value in closed form assuming Laplacian pdf, and show the effect of the correction on image quality. We estimate the Laplacian pdf parameter for each DCT coefficient, and obtain a correction for reconstruction value from the proposed theoretical predictions. The corrected value depends on the Laplacian pdf parameter and the quantization step size Q. The effect of PSNR improvement due to the change in dequantization value is about 0.2 ~ 0.4 dB. We also analyze the reason for the limited improvements.