Unbalanced non-binary tree-structured vector quantizers

Abstract

An established method for developing unbalanced binary tree-structured vector quantizers is greedy growing followed by optimal pruning. These algorithms can be extended to a hybrid binary/quaternary tree structure or to a pure quaternary tree structure. The trade-off of decreased distortion for increased rate is examined for the split into two or four children at each terminal node. The trees employing quaternary splits have smaller memory requirements for the codebook and provide slightly lower mean-squared-error on the test sequence as compared to a binary tree.<>