Pyramid image defogging network based on attention fusion
Caixia Meng, Yuxue Gao, Qin Hou · 2023
Image information captured under foggy conditions is severely degraded, which has an impact on the performance of computer vision tasks. Image defogging has become a critical problem in computer vision. Deep learning-based image defogging methods have made significant progress, but treating different features and pixels equally during deep convolution leads to incomplete defogging and loss of important feature information. Therefore, an end-to-end image defogging network is proposed. The pre-defogging module is first used to generate intermediate results, which contain clear structures, and the feature fusion stage is first learned using channel attention and pixel attention processing to give greater weights to the basic features. Finally the clear image is recovered by compressing the connections with the pyramid full residual module. Extensive experimental results show that the algorithm has good performance on widely used defogging benchmark datasets as well as on real haze images.