Degradation & Restoration: A Low-cost Pipeline for Long-range Single-frame Turbulence Mitigation

ICLR 2026 Conference Submission19612 Authors

Keywords: computer vision, image restoration, turbulence mitigation, long-range observation

Abstract: Long-range turbulence mitigation (TM) remains challenging due to complex spatiotemporal distortions along the imaging path. Current approaches face several limitations in long-range TM: (i) traditional model-based image fusion methods fail to restore dynamic scenes, (ii) learning-based approaches demonstrate either inadequate distortion correction or poor deblurring performance, and (iii) simulators and synthetic training sets inadequately capture the characteristic features of long-range atmospheric turbulence. To achieve optimal restoration with minimal computation, we propose a low-cost single-frame TM pipeline featuring two key innovations: (i) a novel physically-grounded degradation simulator that enables rapid data generation while maintaining fidelity, and (ii) a simple yet effective parallel-training two-stage architecture for sequential distortion removal and deblurring. We demonstrate $4.3\times$ acceleration in degradation simulation and a minimum $2\times$ improvement in training efficiency compared to the baseline. Networks trained on our synthetic data consistently outperform those trained on other SOTA simulations. Our pipeline not only achieves state-of-the-art performance in single-frame TM but also surpasses many multi-frame approaches.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 19612