Semi-Supervised Atmospheric Turbulence Mitigation Based on Hybrid Models

Atmospheric turbulence will degrade the shooting effect of remote imaging equipment in a variety of scenes, because the distortion caused by turbulence involves changes in spatial blur, distortion of geometry, and interference from sensor noise. To mitigate distortion and blurring, this paper proposes a new solution based on the fusion of mathematical interpolation and deep learning. Firstly, TPS (Thin Plate Spline) is used to reduce the distortion of turbulence imaging. Secondly, a simple convolutional neural network containing a lightweight feature extraction residual module was used for deblurring. We have conducted sufficient experiments on different types of turbulence data that the proposed framework can mitigate the motion blur and geometric distortion caused by atmospheric turbulence, thus resulting in a dramatic improvement in visual quality. The proposed network improves the PSNR of image restoration by 0.73-3.50 dB on the algorithm simulation dataset, by 0.23-2.67 dB on the physical simulation dataset, and has a good mitigation effect in real turbulence scenes. Furthermore, an ablation study is conducted to illustrate the enhancements achieved by the various modules in the proposed method.