Rain Streaks Removal for Single Image via Kernel-Guided Convolutional Neural Network.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep learning excels at single image rain removal but often fails due to ignoring motion blur's role in rain streak patterns.
• Existing methods can lead to over- or under-deraining, compromising image quality.

Purpose of the Study:

• To propose a novel rain streaks removal framework that incorporates motion blur information.
• To achieve state-of-the-art performance in single image deraining using a simple yet effective network architecture.

Main Methods:

• A kernel-guided convolutional neural network (KGCNN) framework is proposed, comprising three steps.
• A parameter network learns the motion blur kernel from rainy image details.
• The learned kernel is stretched into a degradation map, guiding a ResNet-based deraining network.

Main Results:

• The KGCNN framework demonstrates effectiveness in removing rain streaks.
• Experimental results on synthetic and real data show superior performance compared to existing methods.
• The method successfully preserves image details during the deraining process.

Conclusions:

• The proposed KGCNN framework effectively addresses limitations of previous methods by integrating motion blur information.
• This approach achieves state-of-the-art results in single image rain removal while preserving image details.
• The KGCNN offers a promising direction for robust image deraining.