Deep learning techniques for atmospheric turbulence removal: a review

Abstract

Atmospheric turbulence significantly complicates the interpretation and analysis of images by distorting them, making it hard to classify and track objects within a scene using traditional methods. This distortion arises from unpredictable, spatially varying disturbances, challenging the effectiveness of standard model-based techniques. These methods often become impractical due to their complexity and high memory demands, further complicating the task of restoring scenes affected by atmospheric turbulence. Deep learning approaches offer faster operation and are capable of implementation on small devices. This paper reviews the characteristics of atmospheric turbulence and its impact on acquired imagery. It compares performances of a range of state-of-the-art deep neural networks, including Transformers, SWIN and MAMBA, when used to mitigate spatio-temporal image distortions. Furthermore, this review presents: a list of available datasets; applicable metrics for evaluation of mitigation methods; an exhaustive list of state-of-the-art and historical mitigation methods. Finally, a critical statistical analysis of a range of example models is included. This review provides a roadmap of how datasets and metrics together with currently used and newly developed deep learning methods could be used to develop the next generation of turbulence mitigation techniques.