Recurrent Multi-Frame Deraining: Combining Physics Guidance and Adversarial Learning.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Existing video rain removal methods primarily address rain streaks.
• These methods often neglect complex degradations like rain accumulation and real-world data priors.
• Current approaches are predominantly trained on synthetic data, limiting their effectiveness on real-world scenarios.

Purpose of the Study:

• To develop a comprehensive video rain removal method.
• To incorporate complex degradation factors and real-world data into the training process.
• To improve the quality of de-rained videos by addressing limitations of existing techniques.

Main Methods:

• A novel two-stage progressive network for video rain removal.
• Stage 1: Inverse recovery guided by a physics-based rain model to estimate background frames.
• Stage 2: Adversarial learning for refining results, restoring color/illumination, and removing artifacts.

Main Results:

• The proposed method demonstrates superior performance on both synthetic and real-world videos.
• Effectively removes rain streaks, accumulation, and other complex degradation factors.
• Achieves state-of-the-art results compared to existing video rain removal techniques.

Conclusions:

• The developed two-stage method effectively handles complex rain degradation factors.
• Combining physics-based recovery with adversarial learning yields significant improvements.
• The comprehensive rain model and proposed method advance the field of video rain removal.