SIMULATED ANNEALING APPLIED TO OPTIMAL DCT QUANTIZATION

Abstract

We present a method for determining optimum quantization tables for use in image compression systems which conform to the ISO/CCITT standard for Image Compression, also known as the Joint Photographic Experts Group (JPEG) standard1. An algorithm based on simulated annealing2 compresses and decompresses any collection of 8 by 8 pixel blocks, while searching the space of 8 by 8 quantization tables for optimum fidelity according to some chosen measure. A composite cost function maintains a predefined compression ratio while minimizing the RMS error in the decoded image compared to the original. The process can be carried out on a raw image, or greater weight can be given to a selected range of DCT coefficients on the basis of psychophysical considerations3. The results of applying the methods to the intensity (y) components of three JPEG test images are presented. In all cases improved fidelity as measured by RMS error is obtained compared to the quantization table suggested in the JPEG standard. Significantly, the quantization tables obtained for one image most often provide smaller error when applied to other images than does the table suggested with the JPEG standard. An unexpected result when using pre-emphasized images suggests that the psychophysical assumptions underlying the suggested JPEG table may be oversimplified. By applying the method to blocks selected from a variety of images, improved quantization tables can be found for images in general, including color.