Deep Learning for Atmospheric Turbulence Mitigation: Data and Models

Imaging through atmospheric turbulence is a significant challenge in long-range optical systems. Turbulence-induced degradation, modeled as spatially varying blur and geometric distortion, hampers not only visual quality but also the performance of computer vision tasks such as recognition and tracking. Traditional approaches, including optical solutions like adaptive optics and model-based restoration, often fall short in scenarios requiring high throughput, dynamic scene restoration, and generalization across variable conditions.

In recent years, data-driven methods have shown promise in tackling image and video restoration. However, the progress of applying data-driven methods for turbulence mitigation has been hindered by three major limitations: the scarcity of realistic and large-scale training datasets; the gap between synthetic and real-world turbulence; and the computational inefficiencies of existing deep turbulence mitigation architectures. This dissertation addresses these challenges from both a physical modeling and learning-based restoration perspective.

To address the data bottleneck, this dissertation introduces improved simulation pipelines based on Zernike-phase models, modified to support large-scale, wide-variety, and physically grounded turbulence synthesis. These simulators incorporate spatial and temporal correlation modeling, adaptive kernel size, and spectral sampling strategies that significantly enhance the realism and efficiency of data generation. They enable the construction of several benchmark datasets that capture a wide range of turbulence conditions and scene dynamics, providing a solid foundation for training and evaluating deep restoration networks.

Beyond simulation, this dissertation explores optimized neural architectures and training strategies for turbulence mitigation. Our work introduces turbulence-specific network design, progressive restoration techniques, simulator-integrated training, and joint restoration-degradation estimation. Our methods consistently outperform existing approaches on both synthetic and real-world turbulence datasets. Notably, our models have been deployed in operational systems for long-range biometric recognition through turbulence, demonstrating both scientific contribution and practical impact of the proposed techniques.