Rate control using embedded adaptive vector quantization

This paper presents a simple, fast and accurate rate control algorithm using adaptive vector quantization (AVQ). The AVQ perspective embeds learning in a dynamic geometry. More specifically, it estimates the unknown joint probability density function p( x, y). The input-output pair represents a sample from the function f:R/sup n//spl rarr/R/sup p/. The function f maps n-vectors x to p-vectors y. We estimate this function using long training image sequences by varying the quantization parameter (QP) within admissible ranges. We generate the non-linear relationship between rate and sum of absolute difference (SAD) using training vectors. Then we apply the mean operator to the rate for each quantized SAD value to generate three look-up tables storing the rate-distortion function. These tables are used to decide the proper QP values for a given bit budget. Simulation results show the proposed algorithm outperforms the MPEG-4 rate control algorithm in terms of reproduced image quality and coding efficiency while requiring much less complexity. It keeps bit rate accurately.