Image compression using vector quantization and artificial neural networks

Among the artificial neural networks, the Kohonen self-organizing feature maps (KSFM) are used for designing a codebook for vector quantization (VQ). Previous studies with KSFM showed difficulties of coding such as edge representing vectors in a codebook, and management of unused codebook entries (nodes). The authors examine problems with the KSFM and propose another coding technique to overcome such difficulties. One postprocessing approach to overcoming the unused nodes problem and classified versions of the KSFM are presented and compared with the LBG algorithm. Several experimental results are presented with KSFM to achieve coding efficiency. The classified KSFM has an advantage over general VQs, especially in terms of computational complexity.<
>