Image coding by a neural net classification process

Abstract

A self organizing neural network performing learning vector quantization (LVQ) to compress image data is proposed. By using unsupervised learning our LVQ neural model approximates optimal pattern clustering from training images through a memory adaptation process and builds a compression codebook in the synaptic weight matrix. The neural codebook trained by example pictures can be used as a codec to compress and decompress other pictures in a speedy fashion. Special image types such as fingerprints verify this property in our experi ments. Our approach is compared with other recently developed neural VQ models - neural gas growing cell structures and conscious competitive learning - and methodological prem ises are discussed. The performance of our model is also compared with JPEG and wavelet methods. Other advantages of our neural codec model are its low training time high utiliza tion of neurons robust clustering capability and simple computation. Further our model has some filtering effects through special training methods and learning enhancement techniques for obtaining a compact neural codebook to yield both high compression and high picture quality. © 1997 Taylor & Francis Group, LLC.