Artificial neural networks for compression of gray scale images: a benchmark

Abstract

In this paper we present results for an investigation of the use of neural networks for the compression of digital images. The main objective of this investigation is the establishment of a ranking of the performance of neural networks with different architectures and different principles of convergence. The ranking involves backpropagation networks (BPNs), hierarchical back-propagation network (HBPN), adaptive back-propagation network (ABPN), a self-organizing maps (KSOM), hierarchically self-organizing maps (HSOM), radial basis function neural networks (RBF) and a supervised Morphological neural networks (SMNN). For the SMNN, considering that it is a neural network recently introduced, an explanation is presented for use in image compression. Gray scale image of Lena were used as the sample image for all network covered in this research. The best result is compression rate of 195.54 with PSNR = 22.97.