Research On Coded Diffraction Imaging Algorithm Based On Plug And Play Priors And Neural Network

In coded diffraction imaging systems,relying on multiple coded diffraction patterns can reconstruct high quality images.However,in the case where the coded diffraction pattern is less and contaminated by noise,it is difficult to reconstruct a high quality image.This paper utilizes the plug-and-play priors model and neural network to improve the quality of reconstructed images.The specific contents are as follows:Firstly,in the paper,based on the importance of image priors for solving image inverse problems,inherent image priors is introduced into the coded diffraction imaging model to improve the quality of reconstructed images by utilizing the NLM(Non-Local Means)denoising operator.And we proposes a coded diffraction imaging algorithm based on plug-and-play priors.Experimental results show that the algorithm can reconstruct higher quality images with less coded diffraction patterns and robust to Gaussian noise.Secondly,we propose a coded diffraction imaging algorithm based on denoising network proximal operator.Because the convolution network can learn more priors of images,we use a trained convolutional denoising neural network as the proximal operators and integrate it into the framework of the proximal gradient optimization method to optimize the coded diffraction imaging model.The proposed method aims to obtain high quality reconstructed images.Experimental results show that the algorithm can not only reconstruct higher quality images,but also robust to Gaussian noise and Poisson noise.Finally,we utilize the plug-and-play priors model and the deep neural network together to improve the quality of the reconstructed image.We propose a coded diffraction imaging algorithm based on plug-and-play priors and proximal gradient neural network.In the algorithm,the coded diffraction imaging process is divided into two steps.Firstly,the algorithm exploits the BM3D(Block-matching and 3D filtering)denoising operator to construct a coded diffraction imaging model based on plug-and-play priors,and relying on the priors introduced by the BM3 D denoising algorithm to improve the quality of reconstructed images.Then the reconstructed image of the process is used as a preliminary estimation,and we utilize the proximal gradient neural network,which is constructed byresidual network structure and aproximal gradient optimization method framework,to further improve the quality of the preliminary estimated image.Experimental results show that the algorithm can not only reconstruct high quality images,but also has robustness to Gaussian noise.