Research On Digital Reconstruction Technology In Coherent Imaging Based On Deep Learning

The object image reconstruction of coherent light field has been the goal pursued in the imaging field.The phase information contains rich object structure information in coherent light field imaging.However,in most cases,it is difficult to directly record the phase information during the data recording process and cause it to be lost,which greatly limits the digital realization of object image reconstruction.In recent years,deep learning algorithms have developed rapidly and have been widely applied in many fields.Deep learning originates from people's research on artificial intelligence technology,and is an important branch in machine learning.With the continuous development and improvement in deep learning algorithms,deep neural networks have made some progress in solving the object image digital reconstruction in coherent light field imaging.Deep neural networks have carried out extensive application research in coherent light imaging field due to their non-linearity,robustness and good learning ability.In this paper,a large number of acquired optical intensity maps and original objects are used as training samples.The non-linear mathematical mapping relationship between the two kinds of data is learned by deep learning algorithms,and a single optical intensity map is used for driving the trained deep neural network model to realize object image digital reconstruction in coherent light field.This paper introduces the relevant physical scenarios of deep learning exploring the object image digital reconstruction in coherent light field.The numerical method of computational optical imaging is used to generate ideal training and testing dataset.The model structure of the deep neural network plays a vital role in the result of object image digital reconstruction.The method of convolution residual neural network is used to combine the multi-layer convolutional neural network into a block-shaped residual structure which applies a trainable convolutional layer on the skip connection structure of the residual,and then adjusts the training parameters of the neural network,and finally obtains a better object image reconstruction result.The main research work includes: 1.Using different off-axis Fresnel digital holograms and related object images to train deep neural networks,analyzing the nonlinear reconstruction effect of the test set at different diffraction distances under the same waveform recording reference light,and comparing it with traditional frequency filtering and four-step phase shift method.2.Using Zernike phase contrast micro-intensity maps and related object images to train a deep neural network,analyzing the nonlinear reconstruction effect under different phase contrast filters.3.The conclusion,the digital reconstruction technology in coherent imaging based on deep learning requires a large number of samples for training,but it can use a single optical intensity map to quickly digital reconstruct the object image,and can realize phase compensation,which is comparable to the traditional method.