In this talk, we propose a physical perception Deep Adversarial Network to recover turbulence-degraded images, gradually eliminating noise, geometric distortion, and blurring. The network accurately extracts distortion features by sensing the geometric distortion field in the degraded image and imposes stronger constraints during the correction process for better results. The feature extraction process utilizes dense residual selfattention blocks to effectively capture geometric distortions and other degradation effects, focusing on the correlation between local and overall features. By learning micro-local and macro global information, the network enhances its ability to perceive distortions and other degradation effects, ensuring coherence between local and overall restoration.