DAWN+: Wavelet-Based Image Deraining Meets Direction-Aware Attention and Mutual Representation.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Single-image deraining aims to remove precipitation artifacts from rainy images.
• Existing methods often overlook the directional nature of rain streaks, leading to degraded textures.

Purpose of the Study:

• To propose a novel direction-aware attention wavelet network (DAWN+) for effective rain streak removal.
• To improve upon the original DAWN model with enhanced capabilities for image deraining and restoration.

Main Methods:

• Introduced vector decomposition for parameterizing rain distribution (vertical and horizontal components).
• Developed a direction-aware attention module (DAM) using coordinate attention for precise rain removal and texture preservation.
• Implemented composite constraints for optimizing structural coherence, detail fidelity, and chrominance accuracy.
• Enhanced DAWN+ by decoupling diagonal coefficient learning, multi-stage decomposition, and cross-frequency mutual representation.

Main Results:

• DAWN+ demonstrated significant performance gains over DAWN, with an average PSNR increase of 1.17 dB.
• The model achieved competitive results compared to state-of-the-art DRSformer, with a 0.15 dB PSNR gain.
• DAWN+ achieved substantial reductions in model parameters (94.4%) and inference time (95%).
• Experiments across six tasks, including deraining, dehazing, and low-light enhancement, showed the model's portability and reusability.

Conclusions:

• DAWN+ offers a significant advancement in single-image deraining by effectively addressing the directional nature of rain streaks.
• The proposed direction-aware attention and enhanced wavelet techniques lead to superior performance and efficiency.
• DAWN+ shows broad applicability across various image restoration tasks, highlighting its robust and reusable strategies.