Predictive Vector Quantization

The coverage of VQ has focused thus far on the coding of a single vector extracted from a signal, that is, on memoryless VQ where each input vector is coded in a manner that does not depend on past (or future) actions of the encoder or decoder. This vector is typically a set of parameters extracted from a finite segment of a signal or a set of adjacent samples of a signal. The segments are themselves usually blocks of consecutive samples of speech or two-dimensional blocks of pixels in an image. Usually we need to quantize a sequence of vectors where each vector can be assumed to have the same pdf, but the successive vectors may be statistically dependent. Separate use of VQ for each vector does not take this dependence into consideration. Vector coders possessing memory may be more efficient in the sense of providing better performance at given bit rates and complexity by taking advantage of the inter-vector or inter-block dependence or correlation.