Direction and Residual Awareness Curriculum Learning Network for Rain Streaks Removal

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Single-image deraining is challenging due to visual similarity between rain streaks and image edges.
• Existing methods often result in over-smoothed edges or residual rain streaks.

Purpose of the Study:

• To develop an effective single-image rain streak removal method.
• To address the limitations of existing deraining techniques by better distinguishing rain from image features.

Main Methods:

• Proposed a direction and residual awareness network (DRAN) within a curriculum learning framework.
• Developed a direction-aware network leveraging the principal directionality of rain streaks in local patches.
• Introduced a residual-aware block (RAB) inspired by iterative regularization to model image-residual relationships.

Main Results:

• Statistical analysis revealed principal directionality in local rain streaks.
• The DRAN method demonstrated superior ability in differentiating rain streaks from image edges.
• Experiments showed significant visual and quantitative improvements over state-of-the-art methods on real and simulated data.

Conclusions:

• The proposed DRAN method effectively removes rain streaks while preserving image details.
• Curriculum learning facilitates progressive learning of rain streak properties and image layers.
• The approach offers a robust solution for single-image deraining challenges.