PFDN: Pyramid Feature Decoupling Network for Single Image Deraining.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Rain streaks degrade outdoor scene visibility, posing challenges for image restoration.
• Existing deraining methods often struggle to simultaneously remove rain and recover fine details due to conflicting objectives.
• The independence of rain streak and background features is typically overlooked in current approaches.

Purpose of the Study:

• To propose an effective Pyramid Feature Decoupling Network (PFDN) for single image deraining.
• To decouple rain-relevant and rain-irrelevant features for improved deraining and detail recovery.
• To enhance the performance of single image deraining through feature disentanglement.

Main Methods:

• Extracting features using a recurrent pyramid module that separates rain-relevant and rain-irrelevant components.
• Employing a novel rain streak removal network for rain-relevant features.
• Utilizing an attention module to enhance rain-irrelevant features for detail recovery.
• Enforcing causality losses to promote feature decoupling across pyramid layers.

Main Results:

• The proposed PFDN effectively models rain-relevant information within the feature domain.
• The framework demonstrates significant performance improvements over state-of-the-art methods on widely-used deraining benchmarks.
• The method shows superiority in the fully-supervised deraining domain.

Conclusions:

• PFDN successfully addresses the limitations of unified frameworks in image deraining.
• The decoupling of features enables simultaneous rain removal and detail restoration.
• The proposed method offers a robust and effective solution for single image deraining.