Variable-Branch Tree-Structured Vector Quantization

Tree-structured vector quantizers (TSVQ) and their variants have recently been proposed. All trees used are fixed M-ary tree structured, such that the training samples in each node must be artificially divided into a fixed number of clusters. This paper proposes a variable-branch tree-structured vector quantizer (VBTSVQ) based on a genetic algorithm, which searches for the number of child nodes of each splitting node for optimal coding in VBTSVQ. Moreover, one disadvantage of TSVQ is that the searched codeword usually differs from the full searched codeword. Briefly, the searched codeword in TSVQ sometimes is not the closest codeword to the input vector. This paper proposes the multiclassification encoding method to select many classified components to represent each cluster, and the codeword encoded in theVBTSVQis usually the same as that of the full search.VBTSVQ outperforms other TSVQs in the experiments presented here.