Multipath tree-structured vector quantizers

Tree-structured vector quantization (TSVQ) is a popular mean of avoiding the exponential complexity of full-search vector quantizers. We present two new design algorithms for TSVQ in which more than one path can be chosen at each internal node. The two algorithms differ on the way the paths are chosen. In the first algorithm the number of paths is fixed and the encoding is similar to the M-algorithm for delayed decision coders. In the second algorithm, the paths are chosen adaptively at each node, according to a (1 + ε) — nearest neighbor rule. We show the performances of the two algorithms on an AR(1) gaussian process, and observe that the adaptive method is the best one. Those methods allow near full-search performances at a fraction of the complexity cost.