Size-adaptive neural network for image compression

A new kind of neural approach to image compression based on a self-adaptive size masking procedure is presented. The neural network (NN) generalization capability has been proved to be a key-element for their application to image compression. In order to improve this feature, an adaptive approach based on pattern classification by activity measure for training and validation is studied. Because of different regions being characterized by different activity, a variable block-size technique has been adopted, in order to improve quality and compression. Different neural networks with different input masks and hidden number are trained on different activity patterns and used on different activity regions. The results have proved this approach to be able to remarkably improve the compression ratio and the global generalization capability of networks. Several tests on learned and unlearned pictures and comparisons with fixed size NNs and DCT-based (JPEG) approaches are reported