Subsampled turbulence removal network

We present a deep-learning approach to restore sequence of turbulence-distorted video frames from turbulent deformations and space-time varying blurs. Instead requiring massive training sample size in deep networks, we purpose strategy that is based on new data augmentation method model turbulence relatively small dataset. Then introduce subsampled enhance the restoration performance presented GAN model. The contributions paper threefold: first, simple but effective algorithm real life for network; Second, firstly Wasserstein combined with $\ell_1$ cost successful turbulence-corrupted sequence; Third, combine subsampling filter out strongly corrupted generate better quality.