Reproduce “Spatio-Temporal Turbulence Mitigation: A Translational Perspective” on OpenResearch

About this paper

Xingguang Zhang, Nicholas Chimitt, Yiheng Chi, Zhiyuan Mao, Stanley H. Chan

Researchers at Purdue University and Samsung Research America developed DATUM, an efficient deep learning network for spatio-temporal turbulence mitigation, which achieved state-of-the-art performance on synthetic data and superior generalization to real-world conditions by integrating classical turbulence insights and leveraging the new physics-grounded ATSyn dataset.

Key insights

• Integrating classical, physics-grounded turbulence mitigation insights (e.g., frame registration, lucky fusion, tilt/blur decoupling) into a deep learning architecture through specialized components (DAAB, GRUM, MTCSA, Twin Decoder) dramatically enhances performance, efficiency, and generalization.
• Developing a highly realistic, large-scale synthetic dataset (ATSyn) with accurate physical modeling of turbulence degradation is crucial for training deep learning models that generalize effectively to complex real-world conditions.
• A recurrent neural network architecture can achieve both state-of-the-art turbulence mitigation quality and high computational efficiency, making real-time deployment feasible for demanding applications.

Summary by alphaXiv.