Residual channel a priori rain removal algorithm guided by elemental attention mechanism
•Innovative adaptive De-raining Framework: introduces the euclidean distance adaptive Multi-stage residual network (EDAMRN), a groundbreaking adaptive framework that dynamically adjusts recursion levels to optimize rain removal efficiency.•Elemental attention Mechanism: incorporates the element atte...
Saved in:
| Published in: | Optics and lasers in engineering Vol. 187; p. 108876 |
|---|---|
| Main Authors: | , , , |
| Format: | Journal Article |
| Language: | English |
| Published: |
Elsevier Ltd
01.04.2025
|
| Subjects: | |
| ISSN: | 0143-8166 |
| Online Access: | Get full text |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | •Innovative adaptive De-raining Framework: introduces the euclidean distance adaptive Multi-stage residual network (EDAMRN), a groundbreaking adaptive framework that dynamically adjusts recursion levels to optimize rain removal efficiency.•Elemental attention Mechanism: incorporates the element attention module (EAM) into the EDAMRN model, enhancing the model's precision in isolating and removing rain streaks by effectively fusing spatial and channel features.•Residual channel prior and interactive Fusion: utilizes residual channel prior (RCP) within the novel EAMRCPN network to leverage background information for guided rain removal. An interactive fusion module (IFM) merges crucial features to pinpoint rain streaks accurately, significantly improving image clarity.•Validated performance Enhancements: extensive experimental results validate the effectiveness of the element attention module in localizing rain streaks. Ablation studies highlight the critical role of integrated modules in advancing de-raining efficacy.•Practical Implications: by enhancing image clarity in rainy conditions, the algorithm improves the usability of outdoor visual systems, aiding subsequent image processing tasks and applications in adverse weather environments.•This condensed highlight summarizes the paper's significant advancements in single-image de-raining technology, emphasizing its innovative methods and impactful outcomes.
Images captured under rainy weather will be obscured by rain streaks, which leads to serious image degradation. Single image rain removal algorithms mainly study how to extract rain streaks information from rain images with complex distributions, and at the same time recover clean images, which facilitates the subsequent research and utilization of image information. In recent years, the emergence of deep learning algorithms has made single image rain removal more accurate and efficient, and a mainstream deep learning-based single image rain removal algorithm combines recursive ideas to achieve progressive rain removal. To address the inefficiency caused by a fixed number of recursions in rain removal algorithms, this paper proposes an adaptive multi-stage deraining network (EDAMRN). By incorporating a Euclidean distance-based recursive controller, the model dynamically adjusts the number of iterations, thereby improving computational efficiency and enhancing output quality. Secondly aiming at the problems of rain streaks residue and background misrepair in the rain image, this paper introduces the EAM (Element Attention Module) into the EDAMRN model. The Element Attention module extracts spatial and channel features from the input and fuses them into attention-enhanced features. The experimental results show that the Element Attention Module helps the model to localize the rain streaks, thus recovering a clear and natural rain removal image. Thirdly aiming at the problem of insufficient combined a priori knowledge in the rain removal algorithm, this paper constructs the EAMRCPN (Element Attention Mechanism-guided Residual Channel Priori Rain Removal Network). EAMRCPN introduces Residual Channel Priori (RCP) to obtain the residual channel features with background information and to better utilize the residual channel features to guide the network to remove rain, it combines the Interactive Fusion IFM(Module), which fuses the rain image features with the residual channel features, to achieve the network's goal of removing rain. Fusion is performed to achieve accurate localization of rain streaks edges by the network, to output high-quality rain removal images. Ablation experiments confirm the critical role of each module and their synergistic integration within the network. |
|---|---|
| ISSN: | 0143-8166 |
| DOI: | 10.1016/j.optlaseng.2025.108876 |