An improvement on competitive learning neural network by LBG vector quantization

This paper presents an integration of a competitive learning neural network with LBG vector quantization for image compression. While LBG works in an off-line style in which the codebook is not updated until all the training blocks are classified, and basic competitive learning neural networks suffer from an under-utilization problem, a new design is conducted in this paper to exploit these two schemes and to propose a new LBG competitive learning neural network. Experiments carried out on image compression show that the proposed neural network significantly outperforms the conventional LBG algorithm in a number of different settings in terms of reconstructed image quality, compression performance and time consumed