Sampling And Reconstruction Jointly Optimized Compression Reconstruction Network And Its Application In Fluorescence Microscopy Imaging

The single-pixel camera based on compressive sensing theory has found widespread applications in fields such as hyperspectral imaging,lidar,and microscopy.Combining photon counting single-pixel compressive imaging technology with fluorescence microscopy can fully exploit the super sensitivity of photon counting single-pixel imaging,which reduces the cost of fluorescence microscopy.However,traditional compressed sensing image reconstruction algorithms are limited by high time complexity and cumbersome parameter settings.Although the deep learning-based end-to-end reconstruction algorithm can avoid a lot of iterative computational costs,one model can only be used for one sampling rate.Using the traditional sampling reconstruction joint optimization compression reconstruction network,the network structure must be modified for different sampling rates,and the model must be trained repeatedly,resulting in a waste of storage space and a multiplied increase in training time.To address these issues,this study aimed to achieve single-model multi-measurement-rate imaging through joint optimization of sampling and reconstruction networks for compressed reconstruction.The main research contents and achievements are as follows:(1)This paper proposes a compressed reconstruction network OMMI-Net that implements single-model multi-measurement-rate imaging through joint optimization of sampling and reconstruction.A training method for fixing the weights of the sampling layer in stages was introduced,which experimentally showed that OMMINet produces a more efficient sampling matrix.Multiple groups of comparative experiments were designed to study the influence of the hidden state dimension of the network recurrent layer and the number of vertically stacked units on the imaging quality,and better network parameters were obtained.A photon-counting microscopy system was constructed,and the network sampling layer weights were binarized for imaging experiments on the system.Experimental results show that OMMI-Net achieves single-model multi-measurement-rate imaging and provides higher reconstruction accuracy than traditional algorithms.(2)A joint optimization compression reconstruction network COMMI-Net based on circular convolution structure is proposed.Drawing on the model-driven design concept,the gradient descent process is embedded in the circular convolution to accelerate network convergence.Compared with OMMI-Net,this method not only reduces the network parameters,but also improves the imaging quality.The imaging performance of COMMI-Net and its variants COMMI-Net V1 and COMMI-Net V2 are compared by simulation experiments.Finally,the COMMI-Net V2 with the best performance is applied to the microscope imaging system for experiments.The results show that COMMI-Net V2 obtains better reconstruction results than OMMINet and traditional algorithms,which verifies the feasibility of the proposed network structure.(3)A data-driven and model-driven fusion compression reconstruction network MOMMI-Net is proposed,which further improves the imaging quality and reduces the network volume.The effect of network cycle times on imaging quality is analyzed through simulation experiments,and MOMMI-Net is compared with other algorithms.The results show that compared with OMMI-Net,MOMMI-Net reduces the number of parameters by 65%,while achieving an average 3d B peak signal-to-noise ratio improvement on the test set.In addition,the reconstruction accuracy of MOMMI-Net is also better than traditional iterative algorithms and some classic deep learning compression reconstruction algorithms.To verify the proposed network model,a photon counting fluorescence compression microscopy imaging system was built,and an imaging experiment was performed on fluorescent microspheres with a particle size of 20 um.Experimental results show that MOMMI-Net realizes single-model multi-sampling rate imaging,and the microsphere image reconstructed at low sampling rate is clear.The network has certain practical application value.