Research On The Application Of Deep Learning In Digital Holographic Reconstruction

Deep learning has now become one of the hot spots in the scientific research field.Especially in digital holography,deep learning solves many difficult problems in the past,such as self-focusing and aberration compensation.However,the requirements of the deep learning are high.First,it requires a large amount of training data,real distribution of object and a large amount of computing resources.Second,it is difficult to obtain the ideal object distribution information.Third,the noise in the acquisition will reduce the reconstruction accuracy.The thesis focuses on the application of deep learning in the reconstruction of digital holography,and focuses on solving the above three problems in it.This paper introduces the principle of digital holographic reconstruction,and introduces the diffraction reconstruction algorithm in detail,and the two methods of removing the zero-order image and the twin image,the Fourier transform method and the phase shift method.We also introduced methods of eliminating phase distortion and achieving super-resolution.It proposed the use of unsupervised algorithm in machine learning——principal component analysis to solve the problems of phase demodulation,aberration compensation,and accurate frequency shift in digital holographic microscopy under structured illumination,which is verified by experiments.A new end-to-end deep learning framework for unpaired data is proposed.It uses a cyclically consistent network structure to input label images and actual recorded holograms into the network for training.Experiments show that this method has two advantages:(1)The limitation of traditional data pairing is broken,and the label set and the training set can be irrelevant;(2)The training data is greatly reduced,even when the number of label sets is half of the training set,the hologram can be reconstructed well using this network;(3)the method has a strong robustness.When the labels are randomly displaced and rotated during the recording process,our proposed algorithm has higher accuracy than traditional deep learning algorithms.In addition,it can also overcome the phase distortion and defocus aberration caused by the imaging system.A method based on deep learning algorithm to suppress speckle noise in digital holographic reproduced images is proposed.It does not require a noise-free image as a label,and only needs to calculate a pair of images with random noise distribution to obtain a noise reduction model.Therefore,the requirements for the stability of the experimental system are reduced,and the true distribution of objects is not required,which enhances the practicality of the algorithm.Experiments prove that the proposed deep learning algorithm can not only remove noise,but also retain the detailed information of objects to the maximum extent.At the same time,in order to prove the universality of this algorithm,this algorithm is also extended to the field of computational holographic imaging.Experiments show that this algorithm can not only achieve the same effect as training with noise-free images,but also use the same algorithm to suppress Speckle noise for 2D and 3D objects.