Study Of The Application Of BP Neural Network To DPCM System Of Image

Differential Pulse Code Modulation (DPCM), which is the popular method of predictive coding, obtains the widespread application in image compression coding because of its simple algorithm and its easiness to realize by hardware. However, considering its linear predictor, the DPCM system, whose differences between the predictive values and the true values are distributed widely, can not get perfect compression ration and signal-noise ration (SNR) in high fidelity coding.Artifical neural network is an information processing unit to imitate the construction and function of human brain. It is composed of plenty of nonlinear units and the complicated connective relationship among these units. Neural network can realize free nonlinear mapping, can do parallel distribution processing, and has very strongly self-learning capability. It has a broad application prospect in the information engineering and the control engineering.The predictor is the core of the image DPCM system, whose precision will directly affect quality of restored images and compression ratio of the data. BP neural network can realize the free nonlinear mapping by the free precision. Therefore, BP neural network will be allowed to used in the predictor of DPCM system, and realizes nonlinear predict.The optimization design of DPCM system is accomplished in this paper applying neural network. In order to get higher compression ration and SNR, BP neural network as the predictor of DPCM system is adopted in this paper, which shrinks the range of the differences compared with nonlinear DPCM system. Construction of an error saturation prevention function and altering to learning-rate is combined to avoid neural network operating in smooth areas for a long term and enhance the operating speed. Aiming at the weights which connect different layers have large difference in adjusting in the standard BP algorithm, this paper makes them adopt different learning-rate to enhance the influence of the hidden layer in the error-adjusting. It is proved that this method can enhance the convergence speed by simulating the sine function. At the same time, in view of the enormous training samples in BP neural network and the excessively unused error in system design caused by enormous data of the image, this paper carries on the method in which the image is divided into several small parts and each one adopts a different BP neural network. From the simulation experiment by C programming, a conclusion is arrived at. Compared with the linear DPCM system, the DPCM system of image used BP neural network as the predictor can further enhances the compression ratio and the signal-noise ratio and obtains the better restored image.