Projeto e avaliação de dicionários para quantização vetorial de voz e imagem.

Abstract

This work presents techniques for designing codebooks applied to vector quantization (VQ) of speech signals and images. The first technique, referred to as SOA (selforganizing algorithm), is inspired on Kohonen's algorithm. The unsupervised learning algorithm SOA, however, uses a neighborhood paradigm which differs from that of Kohonen for updating the codevectors. The second, referred to as SSC (synaptic space competitive), corresponds to an algorithm that uses competitive learning. The third, referred to as FS-SSC (frequency sensitive SSC), introduces Grossberg's conscience principle on SSC algorithm. The fourth technique, referred to as PCA (as an allusion to principal component analysis), computes the VQ codebooks taking into account the eigenvalues and the eigenvectors (principal components) of the covariance matrix of a speech signal. This work presents results concerning speech and image coding based upon simple (conventional) VQ and based upon wavelet VQ, as well as results regarding speaker recognition based upon parametric VQ. Results show that the algorithms SOA, SSC, FS-SSC and PCA are alternatives to the traditional LBG (Linde-Buzo-Gray) algorithm. The computational complexity of the algorithms SSC and LBG is investigated. Analytical expressions (as a function of the codebook size, the dimension of the codevectors, the number of vectors in the training set and the number of iterations executed for codebook design) are derived for the number of operations (multiplications, divisions, additions, subtractions and comparisons) executed by SSC and LBG. Constraints are obtained under which the SSC algorithm is more efficient than the LBG algorithm in terms of number of operations executed in codebook design. The work also presents a method for reducing the computational complexity of the minimum distortion encoding (MDE) of VQ. The proposed method uses the structured organization of the PCA codebooks for significantly reducing the number of operations executed in the process of determining the nearest neighbor for each source vector to be coded, as well as for reducing the memory requirements for codebook storage.