Research On Image Restoration Technology Based On Dynamic Recurrent RBF Neural Network

In the past dozens of years, domestic and international experts and scholars have carried on extensive and deep research to the degraded image restoration and have proposed a lot of effective restoration algorithms. But these traditional restoration algorithms have respective limitations. Artificial Neural Networks (ANN) can adapt the nonlinear model of image restoration effectively without prior knowledge. And the image restoration technology based ANN can avoid some deficiency of the traditional algorithm, which have become a new research focus for these days.RBF Neural Network (RBFNN) is a high effective neural network, it has the best and universal approximation property, simple structure and fast training speed. After deeply studying RBFNN, in this paper we discuss about a Dynamic Recurrent Radial Basis Function Neural Network (DRRBFNN) Model for image restoration, which combines RBFNN and WLP Network. Then we proposed an AGA-DRRBFNN, which uses Adaptive Genetic Algorithm (AGA) to optimize the centers and widths of the network and avoids the influence of different artificial parameter by manual. The main contributions of this thesis are given as follows:(1) By analyzing the technology of image restoration and mainly studying several neural network model used in image restoration, we understand the future trend and direction of the image restoration technology based ANN.(2) We study the basic principle of RBFNN, include the foundation of the theory, its network structure and the mapping relation, and mainly study its training algorithm.(3) We study a Dynamical Recurrent RBF Neural Network (DRRBFNN) and use a hybrid learning algorithm combined with Nearest Neighbor-clustering Algorithm and Gradient Descent Algorithm to train it. We analyze the width parameterγand the convergence parametersK , which have a direct bearing on the classification accuracy and the convergence speed of the network. Several experimental comparative results indicate its advantage in restoring the image with salt-and-pepper noise ratio.(4) After deeply studying the principle of AGA, we use AGA to automatically optimize the parameter of DRRBFNN, which avoids the influence of different artificial parameter by manual on the network performance. Several experimental comparative results for Wine data, Iris data and image restoration validate its efficiency.