Research On Super-resolution Digital Holographic Microscopy Based On Deep Learning

Digital holographic microscopy is a non-contact,full-field quantitative measurement and unlabeled fast 3D measurement method,which is widely used in microscopic imaging and measurement.Its technical process is mainly divided into two parts,holographic recording process and reconstruction process.In the recording process of holographic images,interference from environmental factors and human factors usually leads to the loss of high-frequency information of digital holograms,resulting in a decline in resolution.The reconstruction process of holograms is mainly divided into phase recovery and autofocusing.At present,the commonly used phase recovery algorithms have the problems of high computational complexity and poor processing capability of complex structures.The requirement of autofocusing processing on prior information limits the reconstruction efficiency and greatly reduces the reconstruction performance.The end-to-end deep learning algorithm can effectively improve the problems in the process of recording and reconstruction of the above digital holograms by learning the mapping relationship between low-resolution holographic images and high-resolution reconstructed images,without the need of prior information to improve the details of the image and achieve the super-resolution of the holograms.However,the structure and parameters of deep neural network have great influence on the quality of reconstruction and training speed.Firstly,this paper optimizes and improves the problems existing in the super-resolution reconstruction algorithm of holograms based on convolutional neural network,such as simple network structure,poor recovery effect of detail information and slow convergence speed.The residual network was proposed to solve problems of gradient disappearance or gradient explosion caused by the deepening of network layers,the experimental and comparative analysis of the global residual and local residual structure is also carried out.On the premise that the training speed does not change in order of magnitude,this paper determines the optimal parameters of network depth through experiments,and improves the overall performance of the network after balancing performance and speed.To further improve the ability to recover the details of the hologram,we propose a residual network model based on sub-pixel convolution,to explore the convergence,training speed and test performance of different structures,and the optimal matching number of localresidual block and sub-pixel convolutional layer were determined.Compared with the network structure in the previous section,the detail recovery ability of reconstructed images was comprehensively improved at the expense of longer training time.The experimental comparison between the algorithm in this paper and the traditional algorithm is performed,and the advantages of the algorithm used in this paper are proved from various indicators.Finally,the feasibility of the proposed network structure is verified by simulation and experimental image test sets.By optimizing the structure and parameters of depth neural network,the recovery of unknown samples has achieved a good super-resolution effect,which provides a new idea for extracting high-quality digital holograms more efficiently in the future.