Attention-Guided Unsupervised Network for Image Deraining

Single image deraining is crucial for subsequent computer vision tasks such as object detection, scene understanding, and video analysis. Traditional deraining methods often require end-to-end learning of specific rain streak features, which suffer from poor generalization and lack an automated learning process. Moreover, rain streaks vary across different environments, obscuring and blurring background information while also introducing changes in illumination and color distortion. To address these issues, we propose an Attention-guided Unsupervised Network for Image Deraining, called AUN-ID. AUN-ID incorpo-rates an Attention-based Rain-mask Extraction Block(AREB) to accurately extract rain streaks from images. These extracted streaks are fused with the original rainy image through a Feature Fusion Block(FFB) to obtain a preliminary derained image. Subsequently, leveraging the constrained transfer learning ability and cyclic structure of CycleGAN, we reconstruct clean no-rain images of better quality.