VARIABLE-BRANCH TREE-STRUCTURED RESIDUAL VECTOR QUANTIZATION

Abstract

The traditional tree-structured residual vector quantizers (TSRVQ) are fixed M-ary tree-structured, such that the training samples in each node must be artificially divided into a fixed number of clusters. However, generating the same number of child nodes for each splitting node when the TSRVQ is designed would be inappropriate. This paper presents a genetic splitting (GS) algorithm for designing the tree-structured codebook. The GS algorithm can automatically find the number of child nodes for each splitting node according to the distortion-rate performance. The variable-branch tree-structured residual vector quantizer (VBTSRVQ) based on the GS algorithm is thus proposed. The main disadvantage of VBTSRVQ is that the storage requirements exponentially increase with the number of codewords. However, sharing the codewords in VBTSRVQ can reduce the storage requirements. Thus, this paper also proposes a genetic merge (GM) algorithm to automatically find the similar codewords for merging, and then the storage requirement is reduced in VBTSRVQ. Therefore, this paper proposes VBTSRVQ using a sharing codebook based on the GM algorithm. Also, the distance-rate encoding method is proposed instead of the traditional encoding method for encoding the input sample in VBTSRVQ. VBTSRVQ outperforms other TSRVQs in the experiments presented here.