Robust photon-level single-pixel imaging through diverse scattering media

Imaging through scattering media holds significant application potential in remote sensing, biomedical diagnostics, and industrial detection. However, conventional imaging systems fail to maintain robustness and generalization across diverse environments. Here, we demonstrate a photon-level single-pixel imaging system that exploits data-domain alignment to overcome this limitation. By coupling a physical preprocessing module with a deep neural network, the system translates scattering-induced degradations from different scattering media into a unified data domain, preserving the essential structure of the optical information. Under natural fog and rain conditions, the proposed method clearly reconstructs the fine target details at a distance of 150 m, demonstrating strong robustness across the scattering medium. With 0.088 photons per pattern per pixel in a single measurement, the 256×256 dynamic imaging is well reconstructed. These results establish a generalizable framework for photon-level imaging in diverse scattering media and highlight its promise for robust optical imaging under extreme atmospheric conditions.