Wavelet-driven multi-frequency signal unlocking network for image deraining

Rain streaks and mist caused by rainfall can significantly degrade image quality, directly hindering downstream tasks. In recent years, wavelet transform-based rain removal methods have advanced due to their ability to capture frequency characteristics, thereby enhancing image structure preservation...

Full description

Saved in:
Bibliographic Details
Published in:Neurocomputing (Amsterdam) Vol. 652; p. 131089
Main Authors: Liu, Jiping, Chen, Hongyu, Cao, Shaohan, Wang, Yong, Zhu, Jun, Feng, Dejun, Xie, Yakun
Format: Journal Article
Language:English
Published: Elsevier B.V 01.11.2025
Subjects:
Feature Interaction
Image deraining
Multi-frequency learning
Wavelet transform
Feature Interaction
Image deraining
Wavelet transform
Multi-frequency learning
ISSN:0925-2312
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Rain streaks and mist caused by rainfall can significantly degrade image quality, directly hindering downstream tasks. In recent years, wavelet transform-based rain removal methods have advanced due to their ability to capture frequency characteristics, thereby enhancing image structure preservation while suppressing high-frequency noise. However, existing approaches often overlook the importance of independently processing rain components, disrupting the intrinsic relationship between rain and the background. To address this limitation, we propose a novel framework called the wavelet-driven multi-frequency signal unlocking (WD-MFSU) network. This network employs a wavelet decomposition-driven multi-element learning (WDMEL) mechanism that integrates frequency signals into an attention mechanism and differential learning through a collaborative processing mode, which includes global, structural, and detail-oriented branches. This approach alleviates artifacts and background distortion during rain removal. Additionally, we introduce the rain-background augmented dynamic interaction (RBADC) module to enhance cross-component feature interaction and capture potential dependencies between rain and the background. By utilizing sequential compound attention and dynamic convolution, this module recalibrates channels for rain and background, facilitating dynamic feature adjustments. Comprehensive evaluations demonstrate that our method excels across multiple rainy datasets, outperforming recently proposed approaches while maintaining lower complexity, thus showcasing its effectiveness and generalizability in real-world rainy scenes and downstream tasks. The source code will be available at https://github.com/ChenHuyoo.
ISSN:0925-2312
DOI:10.1016/j.neucom.2025.131089