TransRA: transformer and residual attention fusion for single remote sensing image dehazing

Haze seriously reduces the quality of optical remote sensing images, resulting in poor performance in many applications, such as remote sensing image change detection and classification. In recent years, deep learning models have achieved convincing performance in image dehazing, which has attracted more and more attention in haze removal of remote sensing images. However, the existing deep learning-based methods are less able to recover the fine details of remote sensing images that suffered from haze, especially the cases of nonhomogeneous haze. In this paper, we propose a two-branch neural network fused with Transformer and residual attention to dehaze a single remote sensing image. Specifically, our upper branch is a U-shaped encoder-decoder architecture, using an efficient multi-head self-attention Transformer for capturing long-range dependencies. The lower branch is an attention stack of residual channels to enhance fitting capability of models and complement fine-detailed features for upper branch. Finally, the features of the two branches are stacked and mapped to the haze-free remote sensing image by fusion block. Extensive experiments demonstrate that our TransRA achieves superior performance against other dehazing competitors both qualitatively and quantitatively.