Rate-distortion analysis of dead-zone plus uniform threshold scalar quantization and its application--part II: two-pass VBR coding for H.264/AVC.

Abstract

In the first part of this paper, we derive a source model describing the relationship between the rate, distortion, and quantization steps of the dead-zone plus uniform threshold scalar quantizers with nearly uniform reconstruction quantizers for generalized Gaussian distribution. This source model consists of rate-quantization, distortion-quantization (D-Q), and distortion-rate (D-R) models. In this part, we first rigorously confirm the accuracy of the proposed source model by comparing the calculated results with the coding data of JM 16.0. Efficient parameter estimation strategies are then developed to better employ this source model in our two-pass rate control method for H.264 variable bit rate coding. Based on our D-Q and D-R models, the proposed method is of high stability, low complexity and is easy to implement. Extensive experiments demonstrate that the proposed method achieves: 1) average peak signal-to-noise ratio variance of only 0.0658 dB, compared to 1.8758 dB of JM 16.0's method, with an average rate control error of 1.95% and 2) significant improvement in smoothing the video quality compared with the latest two-pass rate control method.