(A) study on the image compression using neural network

Abstract

The neural network models can be often applied to solve real world problems, especially when the algorithm is not known, or no parallel algorithm is known. From this point of view, there have already been network models proposed for the image data compression by G.W.Cottrell et al. But their result remains a room for improvement because they used one network for large input image patterns and didn``t consider the generalization capability of neural network. In this thesis, in order to improve the quality of reconstructed image, we proposed multiple subnetwork model which consists of pattern classifier and K subnetworks. The pattern classifier classifies the image patterns into one of the subnetworks. Each subnetwork is a feedforward type, 3-layered network with back propagation learning algorithm. The pattern classifier employes picture information measure as a distance measure for clustering. The main feature of our model lies on the was of grouping image patterns, the number of subnetworks and the number of clustes. In order to obtain an improvement of the reconstructed image, similar image patterns are grouped into another subnetwork. The quality of reconstructed image by grouping was examined experimentally. The effects of learning regarding to the number of subnetworks and cluster are also shown experimentally.