Adaptive Vector Quantization

Abstract

Typically in VQ applications, source vectors are sequentially extracted from a real signal and are individually coded by a memoryless vector quantizer. Usually, these vectors are not independent and often they are also not identically distributed. For example, the vector statistics may vary slowly in time and hence exhibit nonstationary behavior. Since successive vectors are in general not independent, the conditional distribution of one vector given the observation of some neighboring vectors provides much more information about the vector than does its marginal distribution. The more “information” we have about a vector, the more accurately it can be coded for a given allocation of bits. Consequently, improved coding performance becomes possible if the quantizer can somehow adapt in time or space to suit the local statistical character of the vector source by observing, directly or indirectly, the vectors in some neighborhood of the current vector to be coded. A vector quantizer is adaptive if the codebook or the encoding rule is changed in time in order to match observed local statistics of the input sequence. Note that the structurally constrained VQ methods described in Chapter 12 are not adaptive since they focus on techniques for coding of a single vector assuming given statistics for that vector and they make no use of the past or future vectors in a sequence.