Deep Learning Based Long-Range Single-Photon Imaging Reconstruction

Single-photon imaging is an imaging method that uses a single-photon detector to reconstruct the target object under the condition of extremely low photon number.Single-photon detector offers extremely high photon sensitivity and picosecond time resolution,which enables accurate measurement of weak signals,such as those returned from very low light intensity or long distances.Single-photon imaging has a wide range of application scenarios,such as high-precision ground surveying and mapping,un-manned driving,medical imaging under low-light conditions,and also has important applications in national defense.In recent years,single-photon imaging has attracted more and more attention in the field of computer vision.This problem has been deeply explored,however,long-range single-photon imaging is still a great challenge,since only a few signal photons mixed with strong background noise can return from multiple reflectors of the scene due to the divergence of the light beam and the receiver's field of view(FoV),which would bring considerable distortion and blur to the recovered depth map.Long-range single-photon imaging based on traditional algorithms has the disadvantages of poor reconstruction quality and long reconstruction time,so how to improve performance from the perspective of deep learning is very important.This dissertation mainly explores the reconstruction of long-range single-photon imaging based on deep learning.We first study the characteristics of single-photon detectors and the physical model of single-photon imaging,and propose a tailored deep learning algorithm on this basis.The research content mainly includes the first use of deep learning to realize the long-range single-photon imaging reconstruction,and the optimization and expansion based on the characteristics of long-range imaging data.Specifically,the main work and innovations of this dissertation can be summarized as follows:1.We study the characteristics of single-photon detectors and single-photon imag-ing models,focus on analyzing the characteristics and difficulties of long-range single-photon imaging,and generate the corresponding simulation data set for network training.To solve this problem,we propose a long-range single-photon imaging reconstruction framework based on deep learning.The reconstruction results are significantly better than the previous algorithms under different noise levels and different sizes of field of view.The superiority of our algorithm is ver-ified in different real-world scenes from kilometers to tens of kilometers away.2.We study the characteristics of long-range single-photon imaging data,analyze the blur effect caused by the divergence of the field of view,and design a series of tailored data preprocessing algorithms to effectively alleviate the blur effect and significantly improve the quality of reconstruction.To further use the photon intensity information in the raw data,we first use the network to estimate the reflectivity map,and then use this to guide the refinement of the depth map,and finally use the depth map to optimize the recovery of the reflectivity map.