Macroscopic-and-Microscopic Rain Streaks Disentanglement Network for Single-Image Deraining

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Single-image deraining is crucial for image restoration.
• Existing methods struggle to effectively distinguish and remove rain streaks, especially from low-frequency image components and preventing edge blurring.

Purpose of the Study:

• To address the limitations in current single-image deraining techniques.
• To develop a robust method for disentangling rain streaks from clean image content.
• To prevent blurry edges during the deraining process.

Main Methods:

• Proposed a self-supervised network for macroscopic rain streak analysis based on pixel distribution.
• Introduced a supervised network for microscopic rain streak analysis using paired rainy and clean images.
• Developed a self-attentive adversarial restoration network to mitigate edge blurring.
• Integrated these into an end-to-end Macroscopic-and-Microscopic Rain Streaks Disentanglement Network (MMRSD-Net).

Main Results:

• The MMRSD-Net effectively disentangles rain streaks from various image components.
• Experimental results demonstrate superior performance on deraining benchmarks compared to state-of-the-art methods.
• The approach successfully addresses the challenges of distinguishing rain, disentangling from low-frequency pixels, and preventing edge blur.

Conclusions:

• The proposed MMRSD-Net offers a significant advancement in single-image deraining.
• The combined macroscopic and microscopic analysis provides a comprehensive solution for rain streak removal.
• The method achieves high-quality deraining results, outperforming existing approaches.