Learning Rain Location Prior for Nighttime Deraining and Beyond.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Existing deraining methods are insufficient for nighttime conditions due to darkness and complex illumination.
• Nighttime rain presents unique visual challenges that vary significantly by location.

Purpose of the Study:

• To enhance nighttime deraining performance by introducing a novel prior learning mechanism.
• To develop a new dataset specifically for nighttime deraining research.
• To improve the generalizability of deraining models to various adverse weather conditions.

Main Methods:

• Proposed a Rain Location Prior (RLP) learned implicitly from rainy images.
• Introduced a Rain Prior Injection Module (RPIM) utilizing multi-scale attention for efficient feature modulation.
• Developed the GTAV-NightRain dataset, capturing rain streak and non-uniform illumination interactions.

Main Results:

• Achieved a 1.3dB improvement in Peak Signal-to-Noise Ratio (PSNR) over state-of-the-art methods.
• Demonstrated superior generalization on real-world data, including heavy rain and glare.
• Validated component effectiveness through ablation studies and visualized RLP interpretability.

Conclusions:

• The proposed RLP and RPIM significantly advance nighttime deraining capabilities.
• The GTAV-NightRain dataset provides a valuable resource for future research in adverse weather image restoration.
• The method shows promise for other location-dependent degradation tasks like daytime deraining and desnowing.