Rate-Distortion Analysis for H.264/AVC Video Statistics

Abstract

MPEG standards family specify the decoding process and the bit-stream syntaxes allowing research towards the optimizations of the encoding process regarding coding performance improvement and complexity reduction. The purpose of a video encoder for broadcast or storage is to generate the optimal perceptual video quality, or the minimized distortion, under a certain constraint such as storage space or channel bandwidth. In particular, by minimizing the distortion D, the video encoder should optimally compute a set of optimal quantisers to control the output bit-rate for each coding unit to satisfy the allocated bit budget. There are two main approaches to solve the optimal bit allocation problem: Lagrange optimization (Everett, 1963; Ramchandran et al., 1994) and dynamic programming (DP) (Bellman, 2003). The optimal bit allocation was first addressed in (Huang & Schultheiss, 1963) where the Lagrange multiplier approach for R-D analysis in transform coding was used. Further improvements have been reported in (Shoham & Gersho, 1988) for source quantization and coding. However, the Lagrange multiplier method suffer from problems, such as having negative bits and real numbers (Schuster & Katsaggelos, 1997a) and the computational complexity is very high due to the need to determine R-D characteristics of current and future video frames. DP is employed to achieve the minimum overall distortion through a tree or trellis with known quantisers and their R-D characteristics (Forney, 1973; Ortega, 1996; Ramchandran et al., 1994). The total of the required bits and coding distortion depend on the quantization step-size. The rate or distortion versus quantization parameter (Q) curve can be produce by encoding for all the possible quantisers to obtain the bit-rate and the quantization error. In order to know how to select a quantization parameter under a specific constraint, e.g., the target bitbudget or distortion, it is importance to model or estimate the coding bit rate in terms of the quantization parameter, namely rate-quantization (R-Q) functions. Together with distortionquantization (D-Q) functions, R-Q functions characterize the rate-distortion (R-D) behaviour of video encoding, which is the key to obtain an optimum bit allocation. Many R-Q and D-Q functions have been reported in previous studies (Chiang & Zhang, 1997; Ding & Liu, 1996; Hang & J.J. Chen, 1997; ISO/IEC, 1993; ISO/IEC, 1997; ITU-T, 1997; Lin & Ortega, 1998;