Variable rate self organizing neural networks for video compression

This paper describes a design of Kohonen's self-organizing neural networks as learning vector quantizers (LVQ) to compress video images. Both fixed rate and variable rate LVQs have been designed. For fixed rate LVQs, both full search and tree-structured codebooks are designed. Further, this paper describes the design of variable rate LVQs. Variable rate LVQs, structured as unbalanced trees, are found to provide improved performance of up to 3 dB peak SNR over comparable fixed rate LVQs.<>