On the Vector Quantization by Means of Recurrent Neural Networks

A vector quantization scheme with a two-stage neural network coding(NNVQ) is developed, where an encoded vector is approximated by the output of the networks driven by a vector from an excitation codebook. The networks and the codebook are optimized to overcome some difficulties in conventional algorithms. The experimental results show that the NNVQ which employs the recurrent neural networks and the optimized learning algorithm performs the best among the reference versions of the VQ algorithms