Hologram Reconstruction Method Based On Deep Learning

The deep learning technology is introduced to reconstruct the off-axis and on-axis digital holograms based on the properties of digital holography(DH).The complex steps in traditional digital hologram reconstruction processes are eliminated by the method we proposed,to achieve a high-precision end-to-end reconstruction in DH.the characteristics of off-axis and on-axis digital holograms,and the traditional reconstruction methods are introduced.In off-axis digital hologram reconstruction,the deep learning technology can directly implement end-to-end reconstruction through the original Fourier transform,frequency selection,frequency shift,and inverse Fourier transform operations.In on-axis digital hologram reconstruction,deep learning technology can effectively remove the interference of twin images and zero-order images,and also can achieve high-precision reconstruction of different types of targets without recording multiple phase-shift holograms.The main contents of the paper are followed:(1)The difference and characteristics of off-axis and on-axis digital holograms are introduced.The influence of the reconstruction distance and frequency selection area on the reconstruction of off-axis digital holograms is discussed.Experiments have found that the size of the frequency-selected region directly affects the resolution of the reconstruction image.The larger of the frequency-selective region,the more high-frequency information is retained and the resolution of the reconstructed image is higher.The reconstruction distance directly affects the resolution of the reconstructed image.When the reconstruction distance is consistent with the recording distance of the hologram,the resolution of the reconstructed image is the highest.(2)The deep learning structure based on U-Net neural network is introduced,and the training set and test set of off-axis digital holograms are generated according to the characteristics of off-axis digital holography.Through end-to-end data training,the end-to-end reconstruction of off-axis digital holograms based on the U-Net network is realized.The structural similarity of the reconstruction results is better than 90%,eliminating the need for frequency selection and frequency shifting in the traditional reconstruction process.The efficiency of reconstruction is improved.(3)The basic methods and characteristics of the reconstruction process of on-axis digital holography are introduced,and the training set and test set of on-axis digital holography are generated according to the characteristics of coaxial digital holography.The U-Net neural network is trained according to the generated training set,and the trained nerve is tested with the test set,which realizes the end-to-end reconstruction of the on-axis digital hologram.The experimental results show that the reconstruction technology based on deep learning can effectively remove the influence of the zero-order image and the twin image in the reconstructed image.The high-quality reconstruction of a single coaxial hologram is realized.(4)The U-Net neural network is tested experimentally using intensity and phase targets.The results show that the neural network reconstruction method based on U-Net has high reconstruction accuracy for both intensity and phase targets.At the same time,by adding different levels of noise to the generated on-axis digital hologram,the network can still maintain high accuracy,which verifies the robustness of the network to different target types and noise.