A vector quantization neural network to compress still monochromatic images
W. Chang, Hamdy Soliman, Andrew H. Sung · 1994
A self-organizing neural network performing learning vector quantization (LVQ) is proposed to compress image data from still pictures. The advantages of the authors' model are its low training time complexity, high utilization of neurons, robust clustering capability, and simple computation; therefore, a VLSI implementation is highly feasible. By learning with self-supervision, the authors' LVQ neural model finds near-optimal clustering from image data and builds a compression codebook in the weight connections. The compression result is competitive comparing with JPEG and a wavelet method which has previously been developed as a fingerprint image compression standard. In addition to implementing LVQ into effective learning rules, the authors also introduce a neuron replenishment technique and a centroid adaptation at class stabilization method to enhance the codebook construction and to yield high picture fidelity. The authors also experiment on the filtering effect of a signal-to-noise ratio weight adaptation and the convolution effect of training with intersectedly subdivided images.>