Cloudformer-CycleGAN: an efficient cloud removal network integrating residual learning and channel spatial attention mechanism

Abstract Removing clouds from remote sensing images poses a significant challenge in image analysis. Despite the widespread application of deep learning methods in the field of cloud removal, their ability to extract and integrate contextual information still needs to be strengthened, which is crucial for describing complex terrain features in remote sensing images. Therefore, we propose a new cloud removal method, Cloudformer-CycleGAN. In this method, we replace the original CNN module in CycleGAN with a Residual Group (RG) module for significant enhancement. The introduction of the Residual Group (RG) module enables the network to focus more on local features in cloud regions, significantly reducing the problem of residual clouds during cloud removal. Experimental results show that the proposed model achieves improvements in multiple image evaluation metrics, demonstrating superior performance in the issue of cloud residue, with a significant reduction in cloud residue in the images.