Neural networks for fast image compression

Abstract

The image compression system aims at reducing the graphics. number of bits transmitted as well as keeping the fidelity of the image, such that at the receiver, the reconstructed image will have little distortion. The design of an image compression system involves three aspects: compression ratio, image distortion, and processing speed. In this thesis, we design and implement a new compression/decompression system to compress a gray-scale image. The system consists of three-layer feed forward neural networks. The compression part includes the input layer and the intermediate layer, while the decompression part consists of the intermediate layer and the output layer. To gain high quality of the reconstructed image, a set of natural networks instead of one network have been used in the system. Each neural network was trained with some image blocks which have similar characteristics. In order to decrease the time for the learning process of the neural networks to converge, an adaptive back-propagation learning algorithm was adopted. In order to keep the generalization capability of the compression system and decrease the number of neurons in the intermediate layer, a preprocessing element is designed which performs the necessary' processing of the image before the image is encoded. The preprocessor includes a predictor which predicts the current input block and a subtracting element which generates residual vectors based on the predictive values and the input vectors. The final results shows that the reconstructed image which was processed by our proposed scheme had a very good Peak-to-peak Signal to Noise Ratio and high compression ratio. Moreover, the parallel architecture inherent in the architecture of the neural networks makes the compression system process very quickly.