Research On Algorithms Of In-line Digital Holographic Reconstruction With Generative Adversarial Networks

Digital holography(DH)can be regarded as a new holography which came up after modern optical information processing technology used in traditional optical holography.DH has digitalized and discretized the procedures from recording holograms to get reconstruction images.Similar to traditional optical holography,DH can be classified into in-line and off-axis digital holography.Compared with off-axis DH,in-line DH has a simpler recording optical path and a higher utilization of spatial bandwidth for CCD or CMOS photoelectric sensors.However,zero-order image,primary image and twin image overlap in the reconstruction image obtained by in-line digital holography.In view of the emergence of zero-order and twin-image items in reconstruction image,this paper designs and researches an end-to-end in-line digital holographic reconstruction network based on GAN network with eliminating the undesirable zero-order and twin-image terms,which is named GU-Net.The works of this paper are as follows:Firstly,this paper introduces the theoretical basis of in-line DH and some methods to remove zero-order and twin-image terms.Aiming at the problem of in-line digital holographic numerical recording and reconstruction,we introduce three common algorithms for numerical calculation of Fresnel diffraction integral and compare their ability to suppress undersampling through a numerical experiment.The traditional algorithms and methods with deep learning are introduced to get noise-free reconstruction image.The "four-step phase shifting method" of traditional algorithms is described in detail,and the simulation experiment is carried out by MATLAB.Next,to overcome the difficulties of removing the undesirable zero-order and twin-image items,here,enlightened by the achievements of generative adversarial networks(GAN),we proposed a fully convolutional neural network,called GU-Net.GU-Net could get a high-quality reconstruction image from a single-shot in-line digital hologram with eliminating undesirable zero-order and twin-image terms.Finally,in order to test the generality of GU-Net,five different experiments on three data sets are carried out.The three data sets are MNIST,Chinese characters and symbols,and CFPW(Celebrities in Frontal-Profile in the Wild)respectively.The quality of reconstruction images obtained by GU-Net on three datasets is very good.And GU-Net performs well not only in reconstructing holograms but also in robustness.Compared with traditional optical methods,GU-Net could get a better reconstruction image.