Enhanced Multistage Vector Quantization with Constrained Storage

Vector quantization (VQ) techniques have so far had limited application in high-fidelity or high resolution applications due to computational and storage complexity. While multistage VQ (MSVQ) offers very low storage complexity, it has a much higher distortion than tree-structured VQ. We extend multistage VQ (MSVQ) to a more general product code structure that allows a gradual trade-off between distortion and storage while maintaining a fixed computational complexity. In this approach, each stage of encoding selects one of a set of codebooks for that stage according to the outcome of the prior stage. A limited number of codebooks for each stage are optimally designed by a constrained-storage VQ technique and are optimally shared by a large number of statistically distinct residual vectors.