Non-line-of-sight imaging via physics-informed cascade learning.

Non-line-of-sight (NLOS) imaging enables the reconstruction of hidden scenes by analyzing indirectly reflected light, with broad applications in autonomous driving, security monitoring, and rescue operations. However, high-fidelity reconstruction is often hindered by the coupling of single-photon avalanche diode (SPAD) detector noises (e.g., dark count rate and time jitter) and the intrinsic ill-posedness of the underlying inversion problem. We propose a physics-informed cascade learning (PICL) method, a framework that integrates model-driven denoising with self-supervised, physics-informed reconstructions. PICL employs a cascaded architecture comprising two specialized networks: a lightweight, SPAD-specific noise separation network (Net1) to decouple signal from mixed interference, followed by a reconstruction network (Net2) that embeds a differentiable forward physical model. This architecture eliminates the need for NLOS large-scale paired datasets while preserving physical consistency. Validated on both synthetic and experimental datasets, PICL demonstrates superior robustness under low-SNR conditions, achieving performance improvement over the state-of-the-art methods. By bridging hardware constraints with algorithmic adaptability, this work provides a scalable and robust paradigm for high-resolution NLOS imaging.