Dequantization bias for JPEG decompression

Abstract

Standard JPEG decompression reconstructs quantized DCT coefficients to the center for the quantization bin. This fails to exploit the nonuniform distribution of the AC coefficients. Assuming a Laplacian distribution, we derive a maximum likelihood estimate of the Laplacian parameter, based on the quantized coefficients available at the decoder, and use this estimate to optimally bias the reconstruction levels during decompression. As a decoder enhancement, this technique is fully compatible with the JPEG standard and does not modify the JPEG compressed bit stream. Extensive simulations indicate that, at typical compression ratios, biased reconstruction results in modest PSNR improvements-about 0.25 dB or higher-and slight subjective improvements, for little or no computational cost. Furthermore, simulations show that the PSNR improvements are very close (within 0.07 dB) to the best theoretically possible.