Physics enhanced single-photon imaging based on deep learning

Single-photon 3-D imaging can recover depth under extremely low photon counts, strong background noise, and occlusions, but purely data-driven reconstruction models often underuse the physical structure of the measurement process. We present a physics-enhanced deep learning framework that incorporates anisotropic and isotropic total variation (AITV) regularization as an explicit prior for single-photon depth reconstruction. The AITV branch provides structure-preserving noise suppression, while a temporal branch enhances photon evidence near the time of flight. The two cues are fused by cross-attention and refined by SWNet, a hybrid multi-scale backbone that combines local convolutional detail with global contextual modeling. Experiments on simulated data, public real-world measurements, and laboratory time-correlated single-photon counting (TCSPC) data show that, among the reproducible baselines evaluated under a unified protocol, the proposed method improves depth accuracy, structural fidelity, and robustness across different photon budgets and signal-to-background ratios. More importantly, the framework provides a modular way to inject physical priors into learned reconstruction backbones, rather than only improving a single network architecture.