Short-wave infrared single-photon single-pixel imaging using deep learning

Short-wave infrared (SWIR) single-photon imaging is highly attractive for long-range imaging due to its high atmospheric transmittance, single-photon sensitivity, and excellent timing resolution. However, the performance of SWIR single-photon focal-plane arrays (FPAs) remains constrained by complex fabrication processes, high noise, limited pixel resolution, and high cost. Single-pixel imaging (SPI) provides a promising alternative by replacing single-photon FPA with a single bucket detector that is much easier to fabricate. However, the trade-off between imaging quality and acquisition speed, together with the limited experimental demonstrations in the SWIR regime, still restricts its practical deployment. Here, we develop a two-stage deep learning framework for low-sampling-ratio SWIR single-photon SPI. In the first stage, encoding patterns are optimized through a differentiable measurement model to capture the most informative structures of the scene, enabling effective reconstruction under strong undersampling using a pseudoinverse-based estimator. In the second stage, a reconstruction network is trained to further enhance the initial reconstruction results by suppressing noise and recovering image details. Simulations under various sampling ratios and noise levels show that the optimized patterns consistently outperform random, Fourier, and Hadamard masks in terms of PSNR, SSIM, and visual fidelity. A 1550 nm single-photon single-pixel LiDAR experiment further validates the effectiveness of the proposed approach in a realistic SWIR imaging scenario. These results provide a practical pathway for integrating SPI with SWIR single-photon LiDAR for long-range, high-resolution imaging.