Real-time mid-infrared single-pixel imaging using a spinning mask based on dual detection and enhanced U-Net

The mid-infrared (MIR) spectral region is vital for molecular fingerprinting in biomedical imaging, material analysis, and industrial sensing. However, conventional MIR imaging systems are limited by their high cost, low sensitivity, and system complexity. To overcome this limitation, we propose a compact, real-time MIR single-pixel imaging system based on a spinning calcium fluoride annular mask and dual-path detection architecture. The primary infrared path captures modulated intensity signals, while the secondary visible-light path provides a reference to resolve the horizontal movement of the target, effectively addressing the issue of horizontal motion insensitivity in previously proposed spinning-mask systems. For reconstruction, we developed a Physics-Embedded Sampling-Aware U-Net (PESA U-Net) that integrates physical modeling, sampling-adaptive attention, and dual-domain loss constraints. The experimental results demonstrate real-time 64×64 imaging at 60 Hz with a sampling rate of 15 %, yielding enhanced sharpness, structural fidelity, and motion sensitivity compared to conventional approaches. This study provides a low-cost, high-performance solution for applications such as gas leak detection.