Parallel Image Compression

Abstract

This thesis presents an efficient parallel algorithm for image compression based on neural network for both gray levels and color images. Data compression grew to be an impottant issue in relation to storage and transmission of infonnation. Digital images are one of the impotiant aspects due to their high storage and tqmsmission requirements. Literature is rich with many of image compression techniques, which have been introduced to handle this problem.

It is known that the input and output layers in an attificial neural networks are much bigger than the hidden layer. This compression nature has been employed to develop the compression algorithm. The compression ratio depends highly on the neur<tl networks structure. The goal of this work is to build a compression algorithm based on enor back propagation neural networks. First there is a training phase, which used to train the network with some standard images with different shapes and sizes. After training the networks, we get weights that minimize the etTors between the original images and the decompressed images as small as possible. The compression algorithm consists of two main patts, natnely compression algorithm, and decompression algorithm. The compression algorithm works like training algorithm but here we apply the algorithm on a single image with weights taken from the training algorithm. No real compression has been achieved because the output of the hidden layer is real value. True image compression occurs when the hidden layer output is quantized before transmission. The quantized output of the hidden layer is the decompressed image.