Image compression via multiresolution feature-based VQ of Hermite-binomial transform coefficients using Kohonen neural network

A new effective feature-based resolution-based vector quantisation (VQ) method for Hermite binomial transform domain image coding is proposed. Hermite-binomial transform is known to be highly relevant to the human receptive field due to its association to the Gaussian derivative models. However, there are more transform coefficients than the original image samples due to the requirement of overlapped windows, which hindered its application to image coding. In the proposed scheme, we apply several VQ subcodebooks to encode image edge profiles and textures at different resolution levels. Simulation results on image coding showed that a high compression ratio can be obtained with good visual quality. Results have also been compared with that of JPEG image coding.