Frame adaptive vector quantization with neural networks

Vector quantization is already known as a very efficient method when used in image coding schemes. Moreover, its performance can be improved by using adaptive techniques, able to include the local properties (frame by frame) of an image sequence. As previously presented by the authors (Lancini et al., 1991), the CL-TS neural network approach to vector quantization offers a powerful solution both in terms of reconstructed quality and computational complexity. The CL-TS algorithm is used in a codebook replenishment based coding architecture. In particular, innovative (local) codebook dimensions and selection methods of its codewords are investigated. Results show improved performance in terms of objective image quality versus coding rate