On Dead-Zone Plus Uniform Threshold Scalar Quantization

摘要

This paper studies the rate-distortion performance of symmetric scalar quantizers to extend previous work published by the first author. We first provide a theoretical analysis of dead-zone plus uniform threshold quantization (DZ+UTQ) for nearly-uniform-reconstruction quantization (NURQ). The quantization performance is particularly investigated for Generalized Gaussian (such as Laplacian and Gaussian) sources using the squared-error distortion measure. According to the analysis, we note that the rate-distortion optimized quantizer is DZ+UTQ with NURQ for Laplacian sources, and is very similar to a DZ+UTQ for a uniform reconstruction quantizer (URQ). We further provide theoretical analysis of rate-distortion constrained DZ+UTQ with NURQ and URQ to cover other Generalized Gaussian sources and give the rate-distortion performance comparison theoretically between NURQ and URQ for DZ+UTQ. We conclude that a URQ is a near-optimal reconstruction rule for many sources, and that a DZ+UTQ classification rule is an effective classifier for it. URQ can be considered as a sub-optimal case of NURQ; the advantage of URQ being its simpler reconstruction rule. Based on the theoretical findings, a new DZ+UTQ quantization rounding technique for URQ is developed and integrated into recent H.264/AVC reference software to improve its encoding performance. Up to 1.0 dB performance improvement is observed, particularly in the very high bit rate range.