JPEG image compression using quantization table optimization based on perceptual image quality assessment

We consider the use of perceptual image quality assessment for quantization table (QT) optimization for JPEG compression. For evaluating performance, we consider the use of the Structural Similarity Index (SSIM) for evaluating distortion in the compressed images. This leads to the study of rate-SSIM curves that replace the traditional use of rate-distortion curves based on the PSNR.We introduce a multi-objective optimization framework for estimating the best rate-SSIM curves. To estimate globally optimal quantization tables, A stochastic-optimization algorithm based on Simulated Annealing is proposed and its variations are studied. We report results on all methods on the Lena image and results from selected methods on the LIVE image quality assessment database. For the LIVE database, compared to the use of the standard JPEG quantization table at quality factor QF=95, QTs based on the training set give average bitrate reductions of 11.68%, 7.7% and an increase of 2.4%, while the SSIM quality changes from -0.11%,+0.05% and 0.12% respectively. In all cases, the results indicate that all considered methods improved over the use of standard JPEG tables.