Research On Spatially Variant Resolution Single Photon Compressive Imaging Technology Based On Deep Learning

Single-pixel imaging technology based on compressed sensing theory has attracted wide attention due to its unique imaging method,and has been applied in many fields such as biomedical imaging,terahertz imaging and hyperspectral imaging.Using a single photon detector as the detector in the single-pixel camera can extend the classic single-pixel camera to the single photon level,realize single photon counting compressive imaging,and has the advantages of low cost and ultra-high sensitivity.But the single-pixel imaging speed is very slow,which limits its application in real-time scenes.Researching the sampling method to obtain the most effective information to reduce the number of measurements and optimizing the reconstruction algorithm to achieve fast reconstruction are two ways to improve the single-pixel imaging speed.Focusing on these two approaches,this paper proposes a spatially variant resolution single photon compressive imaging scheme based on deep learning,the main content and results are as follows:(1)A iterative shrinkage-thresholding network with spatially variant resolution imaging,which is jointly optimized by sampling and reconstruction is proposed.A trainable variant resolution measurement matrix based on spatially variant resolution imaging and the model of the fully connected layer in the neural network is designed by non-uniform pixel combination.A differentiated loss function based on weighting the mean square errors of different resolution sampling areas is designed.A Spatially Variant Resolution ISTA-Net~+(SVR-ISTA-Net~+)is proposed by combining this sampling reconstruction joint optimization model with Iterative Shrinkage-Thresholding Algorithm Network(ISTA-Net~+).The impact of sampling model,loss function and training measurement matrix on network performance is verified through simulation experiments.The results show that the SVR-ISTA-Net~+proposed in this paper can greatly improve the reconstruction quality of the central area of the image compared to uniformly sampling at the same sampling times,and the improvement effect increases with the increase of the sampling rate,making full use of the sampling information.(2)The SVR-ISTA-Net~+is verified on the single photon compressive imaging system.The training method of the binarized neural network is applied to the sampling layer of the network model,so that the sampling layer can be trained to obtain a binary measurement matrix that can be used in the single photon compressive imaging system.Perform Monte Carlo simulation on the system to verify the feasibility and conduct actual experiments.The results show that the network model can also effectively improve the reconstruction quality of the central area of the image in actual application scenes.(3)A iterative shrinkage-thresholding network with spatially variant resolution imaging,which is jointly optimized by convolutional sampling and reconstruction is proposed.Two types of trainable variable resolution measurement matrices with shared weight and unshared weight based on spatially variant resolution and the model of the convolutional layer in the neural network are designed.A non-uniform random step size setting method for different resolution sampling areas is proposed.Apply this variant resolution convolutional sampling model to SVR-ISTA-Net~+.The simulation experiments verify the influence of convolutional sampling and random step size setting on network performance.The results show that the convolutional sampling model that does not share weights proposed in this paper can greatly reduce the number of sampling layer parameters and save system resources while maintaining high-quality reconstruction of the central area of the image.