Optimal blockwise dependent quantization for stereo image coding

Many stereo image compression algorithms exploit the redundancy between the two images in a stereo pair by using disparity compensated prediction. Thus, research in coding of stereo images has focused mostly on the issue of disparity estimation, with less attention being devoted to the equally important problem of allocating bits among the two images. This bit allocation problem is complicated by the dependencies arising from using a prediction based on the quantized reference images. In this paper, we address the problem of blockwise bit allocation for coding of stereo images. The goal is to select the quantization parameters for each block in the reference and di erence images so as to minimize some averaged distortion measure, while meeting any applicable bit budget constraints. In this paper we show how, given the special characteristics of the disparity eld, one can achieve an optimal solution with reasonable complexity, whereas in similar problems in motion compensated video only approximate solutions are feasible. The key observation is to note that the disparity eld is composed of vectors with only (or predominantly) horizontal components. Thus, unlike the motion compensated case, blockwise dependencies are limited to a single dimension, i.e. the blocks in one line in the target image depend only on blocks in the same line in the reference image. Under these conditions we present algorithms based on dynamic programming that provide the optimal blockwise bit allocation. With our experiments based on a modi ed JPEG coder we show gains over standard, independent, bit allocation techniques. For example, the proposed scheme provides higher PSNR, about 1-2dB compared to constant quantization in the whole frame and 0.2-0.5dB compared to disparity compensation with independent blockwise quantization. We also propose a fast algorithm that provides most of the gain at a fraction of the complexity.