Sub-diffraction-limited single-photon 3D imaging based on domain features extraction network at kilometer-scale distance

Single-photon imaging holds extensive application value, yet its imaging quality is often constrained by low signal-to-background ratios (SBRs). In this paper, we designed a network based on the statistical priors of photons, corresponding to the Gaussian statistical prior along the photon time axis, the Poisson statistical prior of photon counts, the varying receptive field requirements for different SBRs, and the deconvolution for solving the photon point spread function (PSF). We designed multiple modules to handle different photon statistical priors and introduced Selective Kernel Network (SKNet) as the feature extraction unit. This enables our method to accurately restore the true spatial distribution of photons. Compared to other methods, we place greater emphasis on the impact of neighborhood characteristics on photons and feature extraction under low SBR. In addition, we designed FoV Module to do deconvolution for PSF at sub-pixel scanning scenarios for sub-diffraction imaging. Experimental results demonstrate our methods, in particularly, achieved the best implementation results to date at low SBR data, which can achieve denoising while preserving more detail features and our methods also achieved twofold sub-diffraction imaging at a distance of 1.5 km using collected real-world data, thereby showing its immense potential to improve the performance of single-photon imaging systems for kilometer-scale imaging.