Research On Single Photon Compressive Imaging Based On Deep Learning

Single-pixel imaging uses Compressed Sensing(CS)theory and can obtain two-dimensional images by a point detector.The single-pixel imaging has the advantage of high sensitivity,so it is widely used in medical detection,space remote sensing,three-dimensional imaging,spectral imaging and other fields.The single photon compressive imaging scheme formed by the single-pixel imaging and photon counting technology can further improve the imaging sensitivity,and has wide application prospects in the field of extremely weak light imaging.Traditional reconstruction algorithms have high time complexity.In order to reduce the reconstruction time of compressed sensing,the research on using deep neural networks to replace traditional compressed sensing reconstruction algorithms has made major breakthrough-This method has proved to be an effective solution.However,most researchers stayed in the stage of numerical simulation and didn't conduct system experiments.For fast and accurate imaging,this paper focuses on the single-photon compressive imaging system and deep learning reconstruction network.The main research contents and achievements are as follows:(1)A single photon compressive imaging system is developed and the Binary Sampling Reconstruction Network(Bsr2-Net)is designed based on the two dimensional residual network(Res2Net).The Bsr2-Net is trained on MNIST dataset and BSDS300 dataset respectively.The design of the Bsr2-Net draws on the idea of block compressed sensing.When reconstructing a high resolution image,the Bsr2-Net reconstructs all low resolution image blocks,and then realizes the high resolution image reconstruction according to the spatial relationship between the image blocks.(2)The performance of traditional reconstruction algorithms,deep learning reconstruction network that was reported recently and the Bsr2-Net is compared when reconstructing handwritten digital images and natural images through numerical simulation.The three reconstruction algorithms are applied to the single photon compressive imaging system.The experimental results are compared and analyzed.The traditional reconstruction algorithms and deep learning reconstruction networks are applied to the different light levels of the single photon compressive imaging system,and analyze the system light levels requirements of these two reconstruction methods.(3)A generalized sampling rate deep learning reconstruction network G_Bsr2-Net is designed based on Bsr2-Net.The G_Bsr2-Net draws on the reconstruction algorithm of computing ghost imaging(CGI)and realizes the generalization of the sampling rate at the cost of reconstruction accuracy.The G_Bsr2-Net can reconstruct images at different sampling rates,which can avoid the problem of having to retrain the deep learning reconstruction network when sampling rate is adjusted.