A fast search algorithm for vector quantization using L2-norm pyramid of codewords

Codewords Distribution-Based Optimal Combination of Equal-Average Equal-Variance Equal-Norm Nearest Neighbor Fast Search Algorithm for Vector Quantization Encoding

Vector quantization encoding requires expensive time to find the closest codeword through the codebook for each input vector. A fast search algorithm for encoding is proposed in this paper. The multilevel elimination criterion is still derived from the three features (mean, variance, and norm) inequalities constraints, but the order of the three inequalities constraints, instead of the predefined order like other conventional multilevel elimination criterion, is optimized to speed up the encoding procedure. In the proposed algorithm, the elimination criterion at the first level is set to mean inequality constraint because of its narrower search width, and the priority order at the second and/or third level of variance and norm inequalities constraints is optimized based on the codewords distribution and the location of input vector in terms of the considered features. The experimental results demonstrate the effectiveness of the proposed algorithm.

Fast vector quantization algorithms based on nearest partition set search

A fast search method for vector quantization is proposed in this paper. It makes use of the fact that in the generalized Lloyd algorithm (GLA) a vector in a training sequence is either placed in the same minimum distance partition (MDP) as in the previous iteration or in a partition within a very small subset of partitions. The proposed method searches for the MDP for a training vector only in this subset of partitions plus the single previous MDP. As the size of this subset is much smaller than the total number of codevectors, the search process is speeded up significantly. The creation of the subset is essential, as it has a direct effect on the improvement in computation time of the proposed method. The schemes that create the subset efficiently have been proposed. The proposed method generates a codebook identical to that generated using the GLA. It is simple and requires only minor modification of the GLA and a modest amount of additional memory. The experimental results show that the computation time of codebook training was improved by factors from 6.6 to 50.7 and from 5.8 to 70.4 for two test data sets when codebooks of sizes from N = 16 to 2048 were trained. The proposed method was also combined with an earlier published method to further improve the computation time.