Vector Quantization Codebook Design

The technique of Vector Quantization(VQ) is a new development of the Shannon's information theory in the application of the source coding. The vector data, which is made up of a number of scalar data, is quantified in the vector space. Vector Quantization(VQ) is an efficient compression technique, whose prominent advantage is the high compression ratio and simple decoding process. In this paper, the VQ codebook design methods are discussed as the major topic of research. The advantages and shortcomings of the Pairwise Nearest Neighbour(PNN) algorithm, GLA algorithm and LVQ algorithm are analyzed by comparing the distortions and computational based on the simulations of these algorithm.The modified algorithms of GLA algorithm with the initial codebook trained by the PNN algorithm is proposed in this paper. The results of experiment for random data show that: The codebook quality of the GLA algorithm with the initial codebook trained by the PNN algorithm is obviously better than that obtained by conventional GLA algorithm, and the reconstruction distortion of the input vector is reduced with calculating consumption not increased.The modified algorithms of LVQ algorithm is also presented in this paper. In the proposed algorithm, the threshold of update times for the renovation of the vector of codebook is set to avoid the situation occurred that only a few vectors of codebook are renovated in the training procedure. So that, most of the vectors in the codebook are renovated in the modified LVQ algorithm. The simulations indicate that the quality of final codebook is improved efficiently in the modified LVQ algorithm, and the reconstruction distortion of the input vector is also reduced significantly.