The Research On Algorithms Of Image Restoration Based On Neural Network

Image restoration is one of the most important and basic issues in the field of digital image processing, full of great theoretical and practical significance. The destination of image restoration is to recover image to make sure that the processed image as near as possible to the original image.The traditional image restoration algorithms, such as Inverse filter, Wiener filter, Kalman filter, singular value decomposition pseudoinverse filter, maximum entropy method and so on, all need calculate higher-dimensional equations, or wide stationary process in recovering process which are the root reasons of image restoration problem not been widely used. The specialists in the image processing field have fully understood the adavantages of neural network techniques such as the abilities of self-educated, self-adaptiveness, great robustness and parallel computing, then applied a variety of neural network models into the image processing field. Using neural network to solve the degraded image restoration problem is one of the applications.By analyzing the algorithm of image restoration based on Hopfield neural network, this paper researched on a new image restoration method which is introduced into transient chaos and wavelet theory to overcome the shortcoming that Hopfield neural network is easily falling into local minimum value and to improve the signal noise ratio and the visual effect of restored images. Experiment results showed this method effectively.The main contributions of this paper are:1. This paper discussed a method for image restoration using Hopfield neural network. In the process of analysing the network update rules, the image gray level of the pixel can use neuron state-variable group weighted scheme instead of a simple sum of M neuron state variables to reduce time and storage complexity, so as to guatantee good fault-tolerant capability and reduce the overall scale of neural network in the meanwhile. Furthermore, using continuous change of the state variables to replace the original step change is to guatantee network energy converge to a global minimum precisely.2. This paper discussed a method for image restoration using chaotic Hopfield neural network. By introducing chaos to overcome the shortcoming that Hopfield neural network is easily falling into local minimum value. The proposed chaotic neural network has richer and more flexible dynamics than Hopfield neural network, so that it can be expected to have higher ability of searching for globally optimal or near-optimal solutions. And in this way, it improved the convergence performance and robustness of the initial value.3. This paper discussed a method for image restoration using wavelet chaotic neural network. The activation function of original network model is Sigmoid function increasing monotonically. It is not as good as the basis function in the ability of function approximation, even it will produce redundant. By introducing wavelet theory into chaotic Hopfield neural network, the new network will have greater ability of function approximation with the activation function composed of Wavelet and Sigmoid function.