Rate-distortion performance of resolution-constrained quantization combined with lossless coding

抄録

We investigate the theoretical and practical rate-distortion (R-D) performance of resolution-constrained quantization (RCQ) combined with lossless coding (RCQ+). Based on the high-rate theory, the required rate difference between RCQ and RCQ+, at a given mean distortion, is found to be the Kullback-Leibler distance (KLD) between the source probability density function (PDF) and its rateless centroid density function. Thus, the rate reduction in RCQ+ is diminished as vector dimensionality increases or as the source PDF approaches uniform density. In the experiments with Gaussian data, the R-D performance of high-rate derivation is verified. Huffman coding is implemented on top of the conventional RCQ methods such as ITU-T G.711 and G.722.2 speech coders, and the rate reduction with RCQ+ is found to be decreased for lower values of the KLD.