Feedback Network for Compact Thin Cloud Removal.

Haidong Ding,Fengying Xie,Yue Zi,Wei Liao,Xuedong Song

DOI: https://doi.org/10.1109/lgrs.2023.3256416

IF: 5.343

2023-01-01

IEEE Geoscience and Remote Sensing Letters

Abstract:The thin cloud removal (CR) technique has great practical value for the application of remote-sensing images. Existing deep-learning-based methods have attained remarkable achievements. However, most of them neglect the inherent feature correlations in deeper layers due to learning successively. In this letter, we propose a compact thin CR network based on the feedback (FB) mechanism, called CRFB-Net, which leverages the high-level features as FB information to modulate shallow representations. CRFB-Net employs the recurrent architecture to achieve such an FB scheme. Specifically, the restoration process does not terminate after obtaining an output. In this case, the output of intermediate iterations will flow into the next iteration as FB. For better utilization of FB, a multiscale feature fusion block (MFFB) is designed to refine the low-level representations from three scales. Furthermore, we introduce a curriculum learning (CL) strategy to train the CRFB-Net by gradually increasing the complexity of restoration, through which a sharper result is produced step by step. Extensive experiments demonstrate the superiority of our CRFB-Net, outperforming state-of-the-art (SOTA).