Scale-Space Feature Recalibration Network for Single Image Deraining.

Area of Science:

• Computer Vision
• Image Processing

Background:

• Computer vision applications like autonomous driving require clear images, but rain severely degrades image quality.
• Existing single image deraining algorithms often struggle to retain fine image details.

Purpose of the Study:

• To develop an advanced single image deraining network that excels in both rain removal and detail preservation.
• Introduce the Scale-space Feature Recalibration Network (SFR-Net) to address limitations in current deraining techniques.

Main Methods:

• Designed a Scale-space Feature Recalibration Network (SFR-Net) incorporating a Multi-scale Extraction Recalibration Block (MERB).
• Utilized dilated convolutions within MERB for rich multi-scale feature extraction.
• Integrated a Subspace Coordinated Attention Mechanism (SCAM) for feature recalibration and noise reduction.
• Employed dense connections and cross-layer feature fusion to enhance feature utilization and network stability.

Main Results:

• The proposed SFR-Net demonstrated superior performance in removing rain streaks compared to state-of-the-art methods.
• The network effectively preserved intricate image details, a common challenge for existing algorithms.
• Experiments on both synthetic and real-world datasets validated the effectiveness of the proposed approach.

Conclusions:

• SFR-Net offers a significant advancement in single image deraining technology.
• The combination of multi-scale feature extraction and attention mechanisms leads to improved deraining performance.
• The method provides a robust solution for obtaining clear images in rainy conditions for computer vision applications.