Spatiotemporal information fusion for photon-level dynamic imaging

In the photon-starved regime, dynamic reconstructions often exhibit attenuated contrast, loss of fine structure, and temporal flicker, which obstruct high-frame-rate, high-fidelity imaging. We present PhotonST-Net, a spatiotemporal deep reconstruction network that fuses spatial structure with inter-frame dependencies through a frame-difference-aware architecture, thereby enforcing temporal coherence and restoring high-frequency detail at ultra-low photon counts. Operating at a sampling ratio as low as 2^- 9, the system achieves 256 × 256 imaging at 200 frames per second with fewer than 2 photons per pixel. It is robust to noise, occlusion, and non-rigid deformation, and an adaptive separation module enables dual-target reconstruction under scattering interference and overlapping motion. Experiments across diverse dynamic conditions validate the approach, showing improvements of 43.12 dB in PSNR and 0.8756 in SSIM compared with conventional block compressive sensing imaging, with further improvements of 15.27 dB and 0.0118 over U-Net under photon-starved conditions. The method offers a scalable solution for high-speed, photon-limited imaging in remote sensing, transient-event observation, and traffic monitoring.