A Fast Full Search Equivalent Formean - Shape - Gain Vector Quantizersj

Vector quantization is an eecient lossy digital signal compression method. It encodes a k-dimensional vector by a reference in a dictionary { or code-book { of representative vectors called codewords. Mean-shape-gain vector quantizers (MSGVQ) encode separately the mean and the length { or gain { of the vector, using two scalar quantizers. The mean-removed, normalized vector is called the shape, and is encoded by an index in a shape codebook. Finding the best shape codeword requires a maximum correlation search which we show to be equivalent to a nearest-neighbor search on a unit hypersphere. We address this problem, and try to nd fast full-search equivalents. An eecient elimination rule based on a a spherical triangular inequality is derived, and combined with the classical partial distance search. Reference vectors for the application of the inequality are the rst principal components of the codewords distribution. We also present some experimental results, and show that our technique is faster than some other existing elimination rules.