Adaptive rate control using nonlinear regression

This paper presents a simple, fast, and accurate rate-control algorithm using nonlinear regression that plays a central role in estimation theory. We measure a conditional mean by estimating a joint probability density function (PDF) using Parzen's (1962) window. The training data pick up the nonlinear rate-distortion (R-D) relationship between the quantization parameter (QP) and the bits spent for each macroblock depending on the sum of absolute differences (SAD). We increase the accuracy of this joint PDF by clustering the training data depending on the QP values within admissible ranges. This localization helps understand image characteristics more accurately. Then we apply the adaptive vector quantization (AVQ) to simplify the conditional mean estimation of the rate given the SAD and QP values. This information is stored into three look-up tables. They contain the localized R-D function on macroblock basis. We use these tables to find the optimal QP values in least-mean-square sense for a given bit budget of the current frame. Simulation results show that the proposed algorithm outperforms the informative MPEG-4 rate-control algorithm in terms of reproduced image quality and coding efficiency. Our algorithm gives better image quality using much fewer bits. Most of all, it keeps the bit rate very accurately due to the accuracy of the conditional mean estimator that solves the nonlinear R-D function.