Research On Data And Model Hybrid Driven Network For Single-Photon Compression Fluorescence Microscopy Imaging

Applying single-pixel imaging and photon counting technology to fluorescence microscopy imaging system has the advantages of ultra-high sensitivity and low cost.Traditional compression reconstruction algorithms usually require a lot of iterative operations to reconstruct images,which have the problems of high computational complexity and long time consumption.The compression reconstruction networks based on deep learning avoid a lot of iterative operations and obtain better reconstruction effect and faster reconstruction speed.Among them,the compression reconstruction networks based on end-to-end learning use data-driven method to train end-to-end network mapping for fitting the dataset.There are problems that the network structure is not interpretable and a large amount of training data is required.The compression reconstruction networks based on model unfolding use model driven method to unfold traditional iterative algorithms into network model,which has the characteristic of interpretability.But there are problems that the network needs to be accurately modeled and some parameters need to be manually set.To solve the above problems,this paper conducts research on the data and model hybrid driven compression reconstruction network,and explores its application in fluorescence microscopy.The main work and achievements are as follows:(1)The data and model hybrid driven network RSOI-Net based on recurrent structure is designed.RSOI-Net utilizes recurrent neural network to optimize measurement values,enabling the backward transfer of information.The results show that the recurrent structure can improve network performance.By studying the impact of different recurrent structures on network performance,the optimal recurrent structure is obtained.Compared with traditional reconstruction algorithms and compression reconstruction networks,the reconstruction effect of RSOI-Net is better.(2)The data and model hybrid driven network MSPOI-Net based on multi-scale feature is designed.In order to fully utilize the multi-scale feature information of images at the granularity level,proximal mapping blocks with controllable unit weighting and attention mechanism weighting based on multi-scale perception is proposed.The results show that the attention mechanism weighting method achieves better reconstruction effect.The impact of the number of proximal mapping blocks and phase blocks on the reconstruction effect is studied through simulation experiments,and the optimal network parameter settings are obtained.Compared with traditional reconstruction algorithms and compression reconstruction networks,MSPOI-Net has achieved better reconstruction effect and faster convergence speed.In order to verify the proposed network,a microscopy imaging system and a single-photon compression fluorescence microscopy imaging system are built,and the sampling sub-network of MSPOI-Net is binarized and applied to the microscopy imaging system.The results show that MSPOI-Net achieves better reconstruction performance than traditional reconstruction algorithms on microscopy imaging system.The imaging experiment of20μm fluorescent microspheres is carried out on the fluorescence microscopy imaging system.MSPOI-Net realize the image reconstruction,and relatively clear fluorescent microsphere image can be reconstructed at a low measurement rate,which verifies the feasibility of the network.(3)The data and model hybrid driven network FEISTA-Net based on feature space is designed.In order to better complement and fuse feature information in the feature space,feature gradient descent block and feature extraction block based on the feature space are designed,and the impact of feature blocks on reconstruction performance is studied through comparative experiments.The results show that the proposed feature blocks can improve reconstruction performance.The impact of the number of phase blocks on reconstruction performance is studied,and the optimal network parameter settings are obtained.Compared with traditional reconstruction algorithms and compression reconstruction networks,FEISTA-Net achieved better reconstruction effect.Compared with training FEISTA-Net on different scene datasets,the reconstruction effect of multi-scene dataset training FEISTA-Net is better.In the imaging experiment of the microscopy imaging system and the single-photon compression fluorescence microscopy imaging system,FEISTA-Net realize image reconstruction.