Research On Lensless Holographic Imaging Algorithm Based On Deep Learning

Lens-free holographic imaging is a new microscopic imaging technique that uses a coaxial holographic approach to imaging using interference between the object light waves scattered through the sample and the reference light waves passing through the sample.The absence of the lens makes the entire imaging system independent of the spatial bandwidth area,enabling high-resolution imaging with a large field of view.However,the recorded holographic image,when directly back-propagating to recover the amplitude and phase information of the object,creates twin image around the object,which seriously affects the recognition of the sample.In order to solve the twin image problem,this thesis proposes lens-free imaging reconstruction algorithms,combining deep learning methods.The lens-free imaging system is introduced and the physical model of lens-free holographic imaging is studied.The causes and effects of twin image generation are analysed.Since there are few publicly available sample complex-valued field datasets on the Internet,this thesis completes the acquisition of 53 sample complex-valued field data by means of lensless microscopic imaging via translated speckle illumination.Microscopic data from 60 samples were collected by an optical microscope filming system.Holographic data sets and microscopic image data sets were produced.A lens-free hologram reconstruction method based on LHR-Unet++ is investigated,which employs a deep network to filter out twin images in backpropagation.the core of the LHR-Unet++ network is the Unet++ network.In this thesis,the convolution block in it is replaced by the double layer convolution block or the residual convolution block,which deepens the network while speeding up the training of the network.In this thesis,LHR-Unet++ is trained using the holographic dataset,with the hologram as input and the back-propagated sample complex-valued field fed into the network,using the phase-recovered complex-valued field as the network label.In order to compare the reconstruction results of LHR-Unet++,this thesis investigates and reproduces the HIDEF hologram reconstruction network,and analyses and compares the reconstruction results of both subjective and objective aspects.The LHR-GAN algorithm is investigated.the LHR-GAN network abandons the previous training method of the network which requires the sample complex value field to be collected as the label of the network in the form of hologram or interference,the algorithm takes the hologram as the input and feds the sample complex value field into the network after back propagation,in order to avoid using the sample reconstructed complex value field as the label,the sample complex value field output from the network will be forward propagated to get simulated hologram.The network will be trained using the minimisation of the difference between the simulated hologram and the input hologram.To avoid the overall network output being a nontrivial solution,a generative adversarial network is added to the amplitude image of the network output,using the observed sample image of the microscope as the discriminator true label to constrain the amplitude image.The designed LHR-GAN network can be trained without the use of the sample complex-valued field,and the microscope images needed can be obtained using a conventional microscope.The LHR-GAN network is trained and its reconstruction results are compared with those of LHR-Unet++,HIDEF and the microscopic sample images.It is demonstrated that the LHR-GAN network trained in this way is indeed successful in removing the twin images and that the reconstruction results are most similar to the microscopic images of the samples.This scheme can be extended to other hologram reconstruction networks.