Image Restoration And Reconstruction Based On Convolutional Neural Network

Due to the advent of the mobile Internet era and the revolution in medical information,images and videos are becoming more and more popular,and image processing is receiving more and more attention.It has developed into a subject with great prospects.Images are inevitably destroyed during the process of acquisition and transmission,which undoubtedly brings great difficulties to researchers.For different processing purposes,image processing is mainly divided into image enhancement,image restoration,image reconstruction,and image segmentation,etc.Image restoration and reconstruction play a very important role in different fields,but it is essentially an ill-posed problem.Generally,by introducing regularization terms,an approximate solution is able to obtain.Traditionally introduced regularization terms can be based on total variation or edge retention.With the development of convolutional neural networks,it provides new research ideas for image restoration and reconstruction.Based on the theoretical method and convolutional neural network model of the image restoration algorithm,this paper studies the application of the traditional mathematical model and convolutional neural network model on different images.This paper consists of image deblurring,image denoising and image reconstruction:(1)A convolutional neural network for deblurred Poisson image restoration is put forward.The end-to-end supervised training convolutional neural network Tnet-Deb is applied to the deblurring problem of natural images.For images with different blur kernels,Tnet-Deb can show better recovery performance.The experiments in this paper verify the feasibility and effectiveness of the proposed network.(2)Inspired by the traditional variance-stabilizing transformation algorithm,this paper constructs a variance stable transformation network(VST-Net)for denoising Poisson images,and analyzes the denoising performance of joint learning and progressive learning methods respectively.At the same time,the paper also explores the parameters of the network.The rational application of the batch normalization layer shows a strong recovery ability for Poisson denoising.A large number of comparative experiments have verified the effectiveness of this method.(3)Different from the traditional optimization model that utilizes artificially set priors,this paper applies denoising auto-encoding network as prior information in image reconstruction,and is employed for sparse projection X-ray CT image reconstruction.In the network learning stage,three-channel images are selected for learning to obtain high-dimensional prior information.In the reconstruction stage,variable augment and channel averaging techniques are utilized for the reconstruction of single-channel CT images.The proposed robust and enhanced denoising auto-encoding prior(REDAEP)method has achieved good results in both parallel beam and fan beam CT reconstruction.The paper also explores the impact of different norms on performance.To sum up,this paper takes convolutional neural network as the core tool,and by utilizing it as a direct mapping or prior information learning,the algorithms proposed in image restoration and reconstruction all show good results.