The asymptotic closed-loop approach to predictive vector quantizer design with application in video coding

The basic vector quantization (VQ) technique employed in video coding belon gs to the category of predictive vector quantization (PVQ), as it involves quantization of the (motion compensated) frame prediction error. It is well known that the design of PVQ suffers from fundamental difficulties, due to the prediction loop, which have an impact on the convergence and the stabi lity of the design procedure. In this paper we propose an approach to PVQ d esign that enjoys the stability of open-loop design while it ensures ultima te optimization of the closed-loop system. The method is derived for genera l predictive quantization, and we demonstrate it on video compression at lo w bit rates, where it provides substantial improvement over standard open a nd closed loop design techniques. Further, the approach outperforms standar d DCT-based video coding.