Image Compression Based On Artificial Neural Network

Image compression based on artificial neural network provides a novel way for investigation of theory as well as technique in the field of image compression. In this paper, we discuss deeply image compression based on back-propagation(BP)neural network and self-organization feature mapping(SOFM)network. Then, we propose a modified vector quantization(VQ)method aiming at the deficiency of original SOFM VQ, which not only demonstrates the potentials of the proposed methods, but also provides a profound insight into the theory of image compression based on artificial neural network. In this paper, two main parts are presented: image compression based on BP neural network and image compression based on SOFM VQ. The former includes image compression based on single and hierarchical BP neural network. The latter mainly includes original SOFM VQ, edge detection from image based on Hopfield network and edge preservation SOFM VQ. In sum, this paper is developed according to the following hierarchy and research mechanism. (1)BP neural network can provide the ability of data compression directly. So we firstly discuss BP algorithm and explode the mechanism of image compression based on BP neural network. Then we make a study of the key technology in the application and several kinds of learning rules are used to compress the static image. After a series of experiments are executed, we analyze and summarize the relationship between the compression performance and the parameters of BP neural network. This is a main part of this paper. (2)In combination with the correlativity between image blocks, we propose the idea of image compression based on hierarchical BP neural network. The further analysis is conducted in the network model and the nested training algorithm. Finally image compression and image reconstruction are accomplished respectively. This method can achieve higher compression ratio and it is a try as well as one of the innovations. (3)The algorithm and architecture of SOFM neural network and the theory of VQ are discussed. Further discussion is made about the key technology in the application. And the VQ based on SOFM is executed and finally we analyze all kinds of parameters have influence on reconstructed image performance. This is the other main part of this paper. (4)The other innovation is extracting edges from image by Hopfield neural network, which is so different from all traditional methods. Using the parallel structure of Hopfield neural network, computation speed is greatly increased and the edge detection based on contents is accomplished. And the statistical feature and edge feature of subimages are calculated according to the result of edge extraction. (5)Aiming at the fatal difficulties of image compression based on BP neural network and original SOFM VQ——edge distortion of the reconstructed image, the edge feature of image is melt into the process of modifying for the weight of SOFM neural network. So a new SOFM VQ method based on edge preservation is proposed. Some experiments denominate its availability. Besides, the reconstructed image using the above method has the improved edge feature and the better visual effect. The modified method is the main innovation in this paper. Experimental results of various simulations show that the research has great theoretical significance and widespread potential uses in practice.