Image compression using neural network

Abstract

In this thesis a neural network based image compression method is presented. Neural networks offer the potential for providing a novel solution to the problem of data compression by its ability to generate an internal data representation. Our network, which is an application of counter propagation network, accepts a large amount of image data, compresses it for storage or transmission, and subsequently restores it when desired. Specifically, a pixel may be viewed as a vector in a three-dimensional space and an image file is a collection of such vectors. So image compression problem may be viewed as a clustering of those vectors into groups based on their similarities. However, the clustering process should 'be as accurate as possible and should have the ability of generalization. In our network a reconstruction vector is defined for each cluster. When the network is presented a new input vector, the cluster in which the vector lies is first determined, and is then represented by the reproduction vector for that cluster. Thus by using an encoded version of this reproduction vector for storage or transmission in place of the original input vector, considerable savings in storage or transmission bandwidth can be realized, at the expense of some distortion. Performance of the network has been evaluated using some stmdard real world images. It is shown that the development architecture and training algorithm provide high compression ratio and low distortion while maintaining the ability to generalize and is very robust as well.