Neural Network Algorithm For Image Processing Based On Variation Model

This paper focuses on the research and exploration of multiple solving methods for variational models proposed in image processing.More specifically,The mathematical modeling of the two mainstream image processing methods is discussed in detail.Based on the idea that the diffusion direction of the diffusion equation can be controlled parametrically,a forward-backward mixed diffusion variational model is proposed.And new neural network method is used to solve the proposed model except for standard differential method.Here we first solve the proposed model using the traditional differential method.And there are two neural network methods presented,Multi-layer perceptron(MLP)and radial basis function(RBF)neural network.The methods are designed based on that the approximation capability of artificial neural network is outstanding.In the both proposed approach,the back-propagation algorithm is used to train the artifical neural network filter aim at minimizing the appropriate customized error function obtained from the forwardbackward mixed diffusion model.Thus the neural networks could be trained in unsupervised mode.The input of the neural network is the pixel of noised image.And the output is considered as the denoised image.The parameters to be adjusted are the weights in linear layer and the neural centers of the RBF.And we give some optimization strategies for parameters in the nerual network.The considered methods are used to remove noise from different kinds of image,such as binary images,MRI medical images and so on.And the experiment get good results compared other model and the traditional different method especially in binary images.In addition,it provides an example for other variational model to use neural network.And there is also much improvement in the structural design of neural networks.