Physics-guided deep learning for enhanced single-photon single-pixel imaging

Extremely low-light imaging plays a crucial role in various scientific research and industrial applications. However, the noise arising from random detection and inherent limitations of the detector greatly compromises imaging quality. Here, we propose a physics-guided deep learning framework for single-photon single-pixel imaging. A physical noise prior model is introduced to accurately characterize the complex temporal photon detection mechanism, enabling a more reliable strategy for data synthesis. Leveraging the large-scale synthesized dataset, an adaptive network is designed to effectively mitigate multi-scale noise distributions while enhancing information preservation. A series of indoor and outdoor far-field experiments demonstrates robust dataset synthesis and high-quality reconstruction. Compared to the classical single-pixel imaging forward model, the proposed technique yields improvements of 10 dB in the peak signal-to-noise ratio and 0.2 in structural similarity compared to common denoising neural networks. These findings pave the way for practical application of single-pixel imaging in extremely low-light scenes.