Rate control and bit allocations for JPEG transcoding

Abstract

An image transcoder that produces a baseline JPEG file from a baseline JPEG input is developed. The goal is to produce a high quality image while accurately meeting a filesize target and keeping computational complexity-especially the memory usage and number of passes at the input image--low. Building upon the work of He and Mitra, the JPEG transcoder exploits a linear relationship between the number of zero-valued quantized DCT coefficients and the bitrate. Using this relationship and a histogram of coefficients, it is possible to determine an effective way to scale the quantization tables of an image to approach a target filesize. As the image is being transcoded, an intra-image process makes minor corrections, saving more bits as needed throughout the transcoding of the image. This intra-image process decrements specific coefficients, minimizing the change in value (and hence image quality) while maximizing the savings in bitrate. The result is a fast JPEG transcoder that reliably achieves a target filesize and preserves as much image quality as possible. The proposed transcoder and several variations were tested on a set of twenty-nine images that gave a fair representation of typical JPEG photos. The evaluation metric consisted of three parts: first, the accuracy and precision of the output filesize with respect to the target filesize; second, the PSNR of the output image with respect to the original image; and third, the subjective visual image quality.