Image compression with self-organizing networks

Abstract

In this paper we evaluate and compare the performance of self-organizing neural networks applied to the task of image compression. The networks investigated are two-layered architectures with linear neurons, and variants of Hebbian learning rules are used to reduce the dimensionality of the inputs while preserving a maximum of information in the output units. Although in theory all networks considered are effectively equivalent to performing the Karhunen-Loève transform, which is the optimal image compression method in the sense that it allows linear reconstruction of the input information with minimal squared error, the results obtained in practice reveal significant differences between the networks. An experimental study has been conducted to demonstrate these differences and thus some light is shed on the suitability of self-organizing neural networks for image compression, particularly in comparison to more conventional methods.