A theoretical exposition to apply the lamda methodology to vector quantization

Vector quantization is a method, used in the lossy compression of voice and images, which can produce results very near to the theoretical limits; however, its principal disadvantage is that the process of search based its functioning on an algorithm of total search, generating a slow process and of a complexity computational considerable. The present work proposes the combination of two algorithms in the creation of a new vector quantization scheme. First, an associative network is obtained applying a learning algorithm for multivariate data analysis (LAMDA) to a codebook generated by means of the LBG algorithm, the purpose of this network is to establish a relation between the training set and the codebook generated by the LBG algorithm; this associative network is a new codebook (LAMDA-codebook) used by the scheme proposed in this work (VQ-LAMDA). Second, considering the LAMDA-codebook as the central element, we use the classification phase of the LAMDA methodology to obtain a rapid search process; the function of this process is generate the set of the class indexes to which every input vector belongs, completing the vector quantization. Furthermore, it is described how to apply the vector quantization scheme proposed to image compression.