Diffusion Equation Image Denoising Model Based On Deep Learning

In recent years,a series of research results have been obtained in the convolutional neural network for image denoising.Its performance is better than the traditional methods.However,the inherent disadvantages of deep learning,such as uninterpretability and unpredictability,cannot be solved.In addition,there are also some shortcoming in image denoising,such as artificial effect.After decades of development,image denoising method based on partial differential equation has been accepted by many people.It is because the denoising algorithm based on partial differential equation has rich mathematical theory,which can be used to study the suitability and error of the solution.However,the diffusion equation method based on partial differential equation tends to produce "speckle" phenomenon.In view of the above problems,this paper has done the following three works:Firstly,based on the idea that the better the initial value of the diffusion coefficient function is given,the better the denoising effect of the diffusion equation will be.Combining deep learning with partial differential equation,We proposed a DnCNN-based diffusion equation image denoising model.The implementation of the model requires two steps.One is to use the DnCNN training noiseless estimation,the other to use the finite difference method to discretize the diffusion equation to obtain the numerical solution of the model.Secondly,the numerical format of model I only contains the gradient information of the image.Inspired by its discrete form,we proposed a diffusion equation denoising model based on gradient prior estimation.From both theoretical and experimental aspects,this paper confirms the feasibility of obtaining gradient estimates through Res Net.In order to retain more detailed information,this paper improves model II and proposes a hybrid diffusion coefficient model.Finally,our new methods are compared with the traditional denoising methods such as PM,BM3 D,WNNM and DnCNN.The experimental results show that our new methods is superior to those of the other methods.What's more,the comparison of the new model proves that the better the initial value of the diffusion coefficient function is,the better the control of the diffusion equation is.