Researchers propose AI framework for clearer infrared video restoration
New model P2GCL uses physics to enhance blurry infrared footage by 0.2775 dB
A team from Tsinghua University has developed the Physical Prior Guided Cooperative Learning (P2GCL) framework, a novel AI system designed to simultaneously enhance atmospheric turbulence strength estimation and infrared image restoration. The framework operates through a cyclic collaboration between two neural networks: TMNet estimates turbulence strength by outputting the refractive index structure constant (Cn2), while TRNet restores infrared image sequences using Cn2 as input. The restored images are then fed back into TMNet to improve measurement accuracy.
Key to P2GCL's performance are a novel Cn2-guided frequency loss function and physical constraint loss, which align training with optical physics principles. In experiments, the framework achieved the best results to date for both turbulence estimation (0.0156 improvement in Cn2 MAE, 0.1065 increase in R2) and image restoration (0.2775 dB improvement in PSNR). The work was documented in an arXiv preprint but later withdrawn by authors on July 30, 2026.
- P2GCL uses two interconnected models (TMNet for turbulence measurement, TRNet for image restoration) in a cyclic workflow
- Achieves 0.2775 dB PSNR improvement in infrared video restoration and reduces Cn2 MAE by 0.0156
- Introduces physics-aware loss functions (Cn2-guided frequency loss and physical constraint loss)
Why It Matters
Could revolutionize surveillance, astronomy and medical imaging by enhancing real-time infrared video clarity under turbulent conditions