URDS: A Dual-Branch ViTs and CNNs Framework for Unpaired Raindrop and Rain Streak Removal

URDS: A Dual-Branch ViTs and CNNs Framework for Unpaired Raindrop and Rain Streak Removal

Rain streaks and raindrops severely degrade image quality and adversely affect vision-based systems in outdoor environments. Removing such rain artifacts is critical for reliable visual recognition. Although supervised deraining methods achieve strong performance, unpaired data remains a major chall...

Full description

Saved in:
Bibliographic Details
Main Authors: Nianyun Liu, Kebin Sha, Junwei Yan, Xinxi Xie
Format: Article
Language:English
Published: IEEE 2025-01-01
Series:IEEE Access
Subjects:
Image deraining
transformer
deep learning
raindrop removal
attention mechanism
Online Access:https://ieeexplore.ieee.org/document/11082152/
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Rain streaks and raindrops severely degrade image quality and adversely affect vision-based systems in outdoor environments. Removing such rain artifacts is critical for reliable visual recognition. Although supervised deraining methods achieve strong performance, unpaired data remains a major challenge due to the absence of pixel-level supervision. In this study, we propose a novel dual-branch GAN-based framework for unpaired rain detection and removal. For rain detection, we introduce a rain perception block that integrates channel sparse transformer and spatial gated convolution, enhancing feature extraction across channel and spatial dimensions and to generate precise rain masks. For rain removal, we design a rain-attentive block that leverages the detected rain mask as guidance, enabling high-performance deraining in an unsupervised manner. Extensive experiments on benchmark datasets demonstrate that our method achieves state-of-the-art performance in unsupervised rain removal. Specifically, it outperforms existing unsupervised approaches on the RainDS dataset. On Rain200L, Rain200H, and RainDrop, it achieves PSNR/SSIM scores of 34.78 dB/0.9637, 24.85 dB/0.8017, and 29.00 dB/0.9095, respectively, with average improvements of 1.50 dB in PSNR and 0.03 in SSIM. Comprehensive ablation studies further validate the critical contributions of both the Rain Perception Block (RPB) and Rain Attentive Block (RAB) modules to the overall system performance.
ISSN:2169-3536

Similar Items