Image compression by edge pattern learning using multilayer perceptrons

A new method for image compression using several multilayer perceptron (MLP) networks is described. Images to be compressed are divided into small sub-image blocks which are processed independently. The MLPs are used to code residual image blocks formed by subtracting the quantised block mean value from each pixel in the original block. To reduce edge degradation in the reconstructed images, several MLPs are used, each representing a different orientation of edge information. Simulation results are presented which demonstrate the effectiveness of this new technique.