Improved Image Compression Using Backpropagation Networks

Abstract

This paper describes an improved image compression scheme using backpropagation networks. The new scheme is aimed at improving the networks' generalisation capabilities, thereby enabling them to effectively compress a wide range of novel images. The networks operate, and are trained on, residual image blocks, thus eliminating the problem of varying average image intensities highlighted by Cottrell et al. Tabulated experimental results and example reconstructed novel images, for the method of [3] and the new technique, are presented, which demonstrate the improved image compression performance gained using this new technique. 131.