Research On Joint Optimization Network Of Sampling And Reconstruction For Single-photon Compression Imaging

The single-pixel imaging technology based on the compressed sensing theory uses a single-point detector to achieve two-dimensional imaging.Using a single photon detector as the detector in a single pixel camera,the classic single pixel camera can be extended to the single photon level,and single photon counting compressive imaging is realized.Compared with the two-dimensional imaging realized by the area array single photon detector,it has the advantage of low cost.At the same time,the detector in the single pixel imaging system can collect the light intensity of multiple pixels,which can realize the so-called ultra-sensitive imaging.Therefore,single photon counting compressive imaging has important applications in the field of extremely low light imaging such as fluorescence imaging,biomedical imaging,and deep space detection.Based on the principle of traditional compressed sensing,single photon counting compressive imaging requires a longer time for compressed measurement and a larger iterative calculation for image reconstruction.In recent years,deep learning methods have been studied for the reconstruction of compressed sensing images,which not only avoids iterative operations to achieve fast imaging,but also obtains better image reconstruction performance.Most of the deep learning compression and reconstruction networks mainly study how to recover images from the measured values on the algorithm,and are not related to the hardware implementation of the actual imaging system.In this paper,starting from the reality of the single photon counting compressive imaging system,we develop a deep learning compression reconstruction network that is jointly optimized by sampling and reconstruction.The main research contents and results are as follows:(1)Propose a joint optimization method of sampling and reconstruction based on deep learning.In this method,the sampling sub-network and the deep reconstruction sub-network are designed in the convolutional neural network to learn the compression sampling process and the image reconstruction process respectively,and the network parameters are updated according to the loss function to realize the joint optimization of the measurement matrix and the reconstruction algorithm.The final result can be used for single photon counting compressive imaging measurement matrix and corresponding image reconstruction algorithm.(2)Designed a model-driven sampling reconstruction optimization iterative network(Sampling reconstruction optimization iterative network,Sroi-Net).Modify the forward and reverse propagation of the network to binarize the weights of the sampling sub-network,which can be used as a binary measurement matrix for single photon counting compressive imaging systems,and expand the traditional iterative algorithm into a deep reconstruction sub-network for image reconstruction.The comparison experiment between simulation and single photon counting compressive imaging system verifies the performance of the binary measurement matrix obtained by the sampling sub-network and the quality of Sroi-Net image reconstruction.(3)A fast non-imaging recognition method for single photon counting compressive imaging system is proposed.Drawing on the idea of the joint optimization method of sampling and reconstruction,this method does not need to reconstruct the image of the target and then recognize it.It can directly classify and recognize the imaging target based on the count value of the detector output photon in the single photon counting compressive imaging system.The results of simulation and system experiment show that this method can still complete more accurate identification even at extremely low measurement rates.