Image Restoration Of The 4f Optical System Based On Neural Networks

Optical wavelet transform is a real-time signal processing method which has developed in recent years. It combines the advantage of both wavelet transform and optical information processing. It is an effective approach to realize image data compression with the characteristic of parallelism and high-speed. Optical wavelet transform is commonly implemented through an optical 4f system. Since there are many errors in the optical 4f system, the recorded images contain a large number of noises. It is of great importance for image data compression to remove the optical noises.With the deep research of artificial neural networks theory, scholars both from domestic and abroad in image processing field have paid great attention to the advantages of neural network technology, such as the abilities of parallel computing,nonlinear mapping and self-adaptivity. Each artificial neural network model has been widely applied in the image processing field. Restoring image from blurred image is one of its major applications.The main contributions of this paper are:1. The fundamental theory, the research status of the image restoration and the artificial neural networks are introduced.2. A new method for image restoration based on a BP neural network is proposed. Since this method is established on the basis of study and training, the problem of establishing the image degradation model and point spread function is completely avoided.3. The improved image restoration algorithm based on neural networks ensemble is proposed. Through training a number of neural networks and integrating the results, the parameters setting of the network configuration is avoided and the generalization ability of neural networks is improved.4. The problem how to use the structure of the system error information in the image restoration procedure is studied. The whole image is partitioned into many sub-blocks, and then every image block is processed by one network with a restoration method based on the overall treatment. Finally, the output of multiple neural networks has been synthesized into a whole image. The problem that the point spread function is space-variant has been effectively solved.