On the use of image quality estimators for improved JPEG2000 coding

An IQA algorithm that performs well in predicting image quality on current image-quality databases does not guarantee the same success in a coding context. During JPEG2000 encoding, it is possible to generate images with vastly different visual qualities depending on how the distortions are allocated to different frequency bands. To test the efficacy of IQA algorithms for JPEG2000 encoding, we collected subjective rankings of JPEG2000-compressed images in which a fixed amount of distortion was allocated to the subbands in various ways. Human subjects provided quality rankings of the distorted images to determine the best and worst strategies of distortion allocation on a per-image basis. The results of this experiment revealed that, in general, higher quality images were achieved by allocating most of the distortion to the finest scale and the least to the coarse-to-mid scales. We also tested the ability of various IQA algorithms in predicting these subjective rankings. Even though these algorithms have been shown to perform reasonably well on JPEG2000-compressed images in popular image-quality databases (e.g., LIVE, CSIQ), our analysis revealed that these algorithms failed to predict the correct rankings on many of the images. Based on the findings from this analysis and the framework of Most Apparent Distortion (MAD), we present a new IQA algorithm MAD_DWT, designed specifically for JPEG2000 coding which uses local DWT coefficient statistics to predict the perceived distortion due to subband quantization.