| Rate-distortion (R-D) analysis and rate control play a key role in video coding and communication systems by providing the R-D optimized compression performance, assuring the successful network transmission of the coded video data, and achieving the best visual quality at the receiver. In the conventional R-D analysis, the bit rate R and distortion D are considered as functions of the quantization parameter q. That is to say, the source models are developed in the q-domain. These source models either have very high computational complexity or suffer from relatively large estimation and control error. In this dissertation, a new framework for R-D analysis, called ρ-domain analysis, is developed, where R and D are studied as functions of ρ which is the percentage of zeros among the quantized transform coefficients. We observe that in the ρ-domain the R-D functions have unique properties which enable us to model and estimate them accurately and robustly. First, we show that in any typical transform coding systems the rate function in the ρ-domain, denoted by R(ρ), is a linear function. Based on Shannon's source coding theorem, we provide a heuristic justification for the ρ-domain linear rate model. Based on this rate model, a unified rate control algorithm is developed for all standard video coding systems, such as MPEG-2, H.263, and MPEG-4 video coding. Our extensive experimental results show that the proposed rate control algorithm outperforms other algorithms reported in the literature by providing much more accurate and robust rate control and buffer regulation. Within the framework of ρ-domain analysis, a generic distortion model is also developed for transform coding of images and videos. Based on this distortion model and the linear rate model described above, an optimum bit allocation scheme is developed in the ρ-domain. It is applied to the MPEG-4 coding at the object level and to the H.263 coding at the macroblock level. Experimental results are presented to show that the proposed optimum bit allocation significantly improves the coded picture quality.; To estimate the R-D functions without context information, we develop a unified source modeling framework in the ρ-domain by introducing the new concepts of characteristic rate curves and rate curve decomposition. With this framework, the R-D functions of the image/video encoder can be accurately estimated before quantization and coding with very low complexity. Based on the estimated R-D functions, a frame-level rate control algorithm is developed for video coding which outperforms the macroblock-level standard TMN8 rate control algorithm. In addition, an encoder-based rate shape smoothing algorithm is developed to control the video encoder such that the output bit stream has both a smoothed rate shape and a consistent picture quality. This is highly desirable in practical video coding and communications. |
