ADAPTIVE VECTOR QUANTIZATION USING A SELF-DEVELOPMENT NEURAL NETWORK

A novel neural network model, called SPAN (space partition network), is presented. This model differs from most of the currently seen neural networks in that it allows a network to adapt its structure by adding neurons, killing neurons, and modifying the structural relationships between neurons in the network. An adaptive vector quantization source coding system based on SPAN is proposed. The basic idea is to use SPAN as an active codebook that can adapt its structure to follow the source signal statistics. The major advantage of using SPAN as the codebook of a vector quantizer is that SPAN can capture the local context of the source signal space and map onto a lattice structure. A fast codebook searching method utilizing the local context of the lattice is proposed and a novel coding scheme, called the path coding method, to eliminate the correlation buried in the source sequence is introduced. The performance of our proposed coder is compared to an LBG coder on synthesized Gauss-Markov sources. Simulation results show that, without using the path coding method, SPAN yields a similar performance to an LBG coder; however, if the path coding method is used, SPAN displays a much better performance than the LBG for highly correlated signal sources. Because the training method is smooth and incremental, SPAN is suitable as the basis for an adaptive vector quantization system. © 1990 IEEE