Total Variation Image Denoising Using Image Prior Information

Image denoising is the process of continuously reducing noise in images,aiming to find the only solution that most close to the original image when the image is degraded.The addition of the prior term in the tipical denoising model has brought a significant impact than before,therefore the design of the prior term is very important to restore the image.Total variation is now the most mainstream prior term,but it relies too much on the image gradient and lacks universality.It affects the correct judgment of the image information,and it is prone to step effect in denoising.In view of the shortcomings of the prior design in the traditional total variation model,an improvement is proposed.The main research contents of the thesis are as follows:First,this paper expands the gradient of the original prior terms in the horizontal and vertical directions on the original basis,observe edge information from more neighborhood directions of pixels to improve the representation power of prior items on image information.The pre-solving primal-dual algorithm is used to optimize the solution model,and the total variation term is converted into a dual form for solution.Through denoising experiments,the improved algorithm effectively improved the original shortcomings.Secondly,this paper makes further improvements on the newly defined variational prior,and introduces the adaptive fractional operator into the improved variational prior to improve the gradient diffusion of the denoising algorithm.Through denoising experiments,it solves the problem of image fragmentation caused by the total variation prior term,suppresses the step effect and improves the efficiency of the algorithm.Finally,due to the artificially designed prior items are limited in extracting image feature information,the regular term of the objective function in the denoising optimization problem was replaced by the convolutional neural network with the prior information learning ability.The performance of the mean square error loss function often used by networks is poor,and the improved adaptive fractional total variational expression is added to the loss function of the convolutional neural network.Through denoising experiments,the application effect of this algorithm in image denoising has been significantly improved.