STRRNet: Semantics-guided Two-stage Raindrop Removal Network

Qiyu Rong, Hongyuan Jing, Mengmeng Zhang, Jinlong Li, Mengfei Han; Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR) Workshops, 2025, pp. 1370-1378

In this paper, we propose a Semantic-guided Two-stage Raindrop Removal Network (STRRNet) for multi-scene raindrop removal. The multi-scene raindrop removal includes four scenarios: daytime background-focused images, daytime raindrop-focused images, nighttime background-focused images, and nighttime raindrop-focused images. This task poses three major challenges: how to achieve effective raindrop removal under both day and night conditions, how to recover blurred backgrounds while removing raindrops, and how to handle densely distributed raindrops. To address these challenges, we introduce a Semantic Guidance Module (SGM) that integrates semantic information with image features to enable more fine-grained restoration. In addition, we design a Background Restoration Module (BRM) that performs secondary restoration on the initially derained output, aiming to recover blurred background regions. Finally, we apply median fusion to multiple frames of the same scene, which effectively eliminates raindrop residues and artifacts. The fused images are treated as pseudo ground truth to implicitly augment the training data, enabling semi-supervised fine-tuning of the model. This strategy significantly enhances the model's performance on unlabeled images. The proposed STRRNet achieves state-of-the-art performance and won First Place in the NTIRE 2025 The First Challenge on Day and Night Raindrop Removal for Dual-Focused Images.