A Frequency-Sensitive Competitive Learning Networks with Hadamard Transform Applied to Color Image Compression

The neural network is useful for data compression if the connection weights are chosen properly. In this paper, a Modified Frequency-Sensitive Competitive Learning (MFSCL) network with Hadamard transform based on Vector Quantization (VQ) for color image compression is presented. The goal is apply a spread-unsupervised scheme based on the modified competitive learning networks so that on-line learning and parallel implementation for color image compression based on VQ in Hadamard Transform (HT) domain are feasible. In the MFSCL network, each output neuron (codevector) incorporates a count of the number of times it has been the winner. The distortion measure used to determine the winner is updated to include the count number. The color image information transformed by HT operation was separated into RGB 3-plane DC value and AC coefficients. Then the AC coefficients for each plane were trained using the proposed MFSCL method to generate the VQ codebook. The experimental results show that promising codebooks can be obtained using the presented spread MFSCL network for color image compression in the transform domains.