Study On Digital Holographic Microscopic Imaging Based On Deep Learning

Digital holographic microscopic imaging technology is an organic combination of information optics,optoelectronics technology and computer science and technology,which has the advantages of no damage,full field of view,non-contact,flexible processing,convenient storage,and can obtain three-dimensional quantitative amplitude and phase distribution.In recent years,with the rapid development of deep learning,neural networks based on deep learning emerge in an endless stream,and many previously unsolvable problems in fields can be made breakthroughs through deep learning methods.Its applications in the field of digital holography,such as unwrapping,self-focusing and aberration compensation,are becoming more and more extensive,which opens up a new research direction in the field of microscopic imaging.Through theoretical analysis of deep learning network structure and experimental verification of simulation diagrams and real samples,this thesis studies the improvement of signal-to-noise ratio,phase unwrapping,super-resolution and self-focusing reconstruction,etc.The main work contents include:First,In this thesis,a method based on AU-Net to improve the SNR of hologram fringes is proposed to suppress speckle noise and obtain high quality holograms.The mapping relationship between low resolution hologram and high resolution hologram is learned to obtain the noise reduction model.The proposed algorithm reduces the requirements for the stability of the experimental system.Experimental results show that the proposed deep learning algorithm can not only reduce the noise of the hologram,but also suppress the speckle noise in the reconstructed phase map,and display the details of the experimental sample to the maximum extent.Second,In this thesis,an improved U-Net phase unwrapping method is proposed.The SE channel attention module is added after the residual block,and the depth-separable convolution is used to replace part of the traditional convolution.A large number of simulated phase maps randomly generated by the matrix are used as data sets for training to achieve the purpose of unwrapping the real hologram phase.This method unwraps through different experimental samples,and compared with the results of other unwrapping algorithms,the results of the experimental results show that the proposed method in this thesis sample the edge smooth,flat background,experimental test sets and the results of discrete cosine solution package structural similarity index from 0.932 to 0.973 on average,peak signal-to-noise ratio by an average of 21.60 to 29.18.Finally,This thesis proposes an end-to-end reconstruction method of multi-scale holograms based on deep learning.On the one hand,it can realize super-resolution reconstruction,and on the other hand,it can realize self-focusing reconstruction of holograms with different defocus distances.The improved network models GUE-Net and CUE-Net have less parameters,less computation and faster speed.The performance of the proposed network was verified with human red blood cells and chicken blood cells.Compared with the traditional method,the background is smoother,the effect of speckle noise can be suppressed,and the reconstruction speed is faster.Using the trained 256×256 pixel size model to directly test the hologram with large field of view,multi-scale holograms can be reconstructed,and the phase distortion and defocus aberration can be compensated.By optimizing the structure and parameters of the deep neural network,the superresolution and self-focusing reconstruction of multi-scale biological samples with large field of view is realized,and the quality and speed of digital holographic reconstruction images are improved.