Atmospheric turbulence introduces complex spatial and temporal distortions that can severely degrade images in applications such as ground-based astronomy and long-range surveillance. In a new paper in the Journal of Electronic Imaging, Juan S. Carrillo and Jérôme Gilles compare two deep learning approaches—Generative Adversarial Networks (GANs) and Transformer-based models—for mitigating these effects, using the SOTIS dataset, which contains multi-frame sequences with ground truth under varying turbulence conditions.

The central finding is a clear division of labor. GAN-based models performed better in severely degraded, low-quality situations, while Transformer-based models achieved better results in moderate to high-quality settings, particularly when correcting geometric distortions. This suggests that the best architecture depends on the imaging environment, rather than a single model being universally superior.

The authors also introduced several enhancements to the Turbulence Mitigation Transformer (TMT): a refined training strategy, an augmented input design that includes a temporally averaged frame as prior structural information, and a modified loss function incorporating the Structural Similarity Index (SSIM). These changes produced consistent performance gains across turbulence conditions, offering practical guidance for selecting models in real-world imaging systems.