Image compression using a neural network with back-propagation

We present an image compression system based on a feed-forward neural network with one hidden layer, trained with the Error Back-Propagation algorithm. The training is done in batching mode, two neuron activation functions (sigmoidal and linear with saturation) are compared and the influence of several parameters on network behaviour and system performance is analyzed. The usefulness of a few simple techniques of image pre-processing is also tested. The simulations are performed on a INTEL i860 superscalar processor, using a large data-base of images, grouped into various training and testing sets. The system provides good compression quality and generalization capability; we obtain compression ratios up to 64 with an average NMSE nearly always less than 1% for the learned images and only slightly worse for the unlearned ones.