Physics-informed ultra-low photon level single-pixel imaging

Low-illumination imaging is essential for minimizing photodamage, enabling quantum and low-light sensing, deep-tissue imaging under power-constrained conditions. But high-quality, ultra-low-photon-level imaging remains a significant challenge due to the complexity of hardware and the presence of noise interference, especially in non-visible wavebands. Single-pixel imaging (SPI) has attracted considerable interest across a wide spectrum because of the low cost, broadband and highly sensitive response. Here, we propose a physics-informed neural network (PINN) enhanced SPI to achieve image reconstruction at the single-photon level using a single-photon avalanche diode (SPAD) detector. PINN enables highquality image reconstruction under ultra-low photon irradiation without requiring a large dataset for pre-training. Experimental results confirm that PINN-SPI can achieve high-quality image reconstruction under the illumination levels ∼0.02 photons per pixel, with a low sampling ratio of only 2%. This method offers an alternative approach for high-fidelity imaging under extremely weak illumination across X-ray, infrared, and terahertz wavebands, enabling non-invasive detection, improved signal-to-noise performance, and broader applicability in biomedical imaging, LiDAR, and nondestructive testing.