Adaptive vector quantization for fixed bit-rate videocoding

Abstract

Abstract In this paper a new algorithm is introduced and discussed for self-adaptive vector quantization of video sequences. The structure of the vector quantizer and the mechanisms for its construction and updating are based on a binary tree whose topology and content adaptively track the varying statistics of the incoming data. Some tree branches are extended more than others depending on the time varying distribution of the input vectors, to guarantee an acceptable reconstruction quality, not far from the target performances of a full search approach. The resulting bit-rate stems from several terms carrying information for both frame reconstruction and codebook updating. An effective policy for codebit allocation has been devised and tested providing very good results within fixed bit-rate applications. Exhaustive simulations have been carried out on standard sequences at a rate of 30 frames/sec with progressive transmission. An average reconstruction quality of 34.5 dB in the peak signal-to-noise ratio has been achieved for the luminance component at a constant bit-rate of 512 kbit/sec corresponding to a compression of 30:1 (0.26 bit/pixel).