Image compression using dynamic clustering and neural network

Abstract

In this thesis dynamic clustering and neural network techniques have been used for image compression. First we cluster the whole image dynamically up to an expected Peak Signal Noise Ratio (PSNR) and Mean Square Error (MSE). Thus number of clusters depends on the similarity threshold, target PSNR and MSE. Each cluster has its own cluster prototype and cluster identification number. Naturally, each pixel of the image is represented by its cluster identification number. According to the locality property of an image many consecutive pixels happen to be the members of the same cluster. So, a structure consisting of cluster number and repeat count to represent a number of consecutive pixels have been proposed for better compression ratio. Then the cluster number part is further compressed using a neural network. Here the co-ordinates of the pixels are inputs and the corresponding cluster numbers are the outputs of the neural network. Cumulative cluster number is used instead of ordinary cluster number to make the neural network small for larger compression ratio.