Video-rate terahertz single-pixel imaging via physics-enhanced deep learning and VCSEL array modulation

Terahertz (THz) single-pixel imaging has emerged as a significant research direction due to its cost-effectiveness, low system complexity, and high scalability. However, practical implementation is hindered by a critical trade-off between imaging speed and quality, stemming from the combined limitations of modulator response speeds and reconstruction algorithms. Addressing this challenge, we propose a physics-enhanced deep learning framework for video-rate THz single-pixel imaging. By utilizing a high-speed vertical-cavity surface-emitting laser array to pump a high-resistivity silicon wafer, we achieve high-speed THz modulation and remove the refresh-rate limitation associated with commonly used DMD-based optical pattern generation. To bridge the gap between hardware addressing constraints and the requirements for complex encoding masks, a low-rank matrix synthesis strategy is integrated into the neural network training to jointly optimize mask distribution and image reconstruction. Both simulation and experimental results demonstrate that this synergistic mechanism outperforms pure data-driven algorithms and traditional basis scanning methods, enabling real-time THz video capture at 50 f/s with a resolution of 32 × 32 pixels. Even at a 25% sampling ratio, the system achieves high-fidelity reconstruction, effectively balancing speed and quality. This hardware–software co-designed approach paves the way for real-time THz applications in industrial and scientific research while establishing a new paradigm for high-speed dynamic imaging across other spectral regimes.