MAE-ADHRN: Adaptive degradation high-resolution reconstruction model applied to remote sensing images
Lixin Pu, Fangjie Dong, Jun Zhang, C. Li, Yuxin Pu, Jipeng Fan, Mingjie He, Zhongjun Gao · 2024
Abstract: The reconstruction of complex degradation of remote sensing images is always a hot issue. In practical application, there are many reasons for complex degradation. In order to solve the complex remote sensing image degradation problem, MAE self-supervised model is introduced to form prior knowledge. Secondly, a new remote sensing image reconstruction network and edge focus module are designed. The edge focus module learns the weight of edge position by calculating and extracting gradient information, so that the reconstructed model can pay attention to the edge information of remote sensing image. Finally, the two-stage training strategy is introduced. Experimental results show that the MAE-ADHRN (MAE +Ours) model proposed in this paper achieves better results than other models based on complex image degradation. At the same time, after the introduction of edge attention module and high-resolution remote sensing image, the edge information is clear. The PSNR and SSIM of MAE+Ours are 3.899 and 3.024 higher than YOLO+Ours and MoCo+Ours, respectively. LPIPS decreased by 0.0513 and 0.0341, respectively. Ablation results show that compared with None-EI and None-Edge attention modules, PSNR and SSIM of edge attention modules are increased by 2.648 and 0.397, respectively, and LPIPS are decreased by 0.0268 and 0.0037, respectively.