Towards Unified Deep Image Deraining: A Survey and a New Benchmark

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Significant advancements in image deraining have emerged from improved image priors and deep learning.
• Existing deraining methods lack standardized evaluation settings, hindering fair comparison and assessment of practical utility.
• Previous surveys on image deraining have not focused on unifying evaluation criteria.

Purpose of the Study:

• To provide a comprehensive review of current image deraining techniques.
• To establish a unified evaluation setting for assessing deraining method performance.
• To introduce a new benchmark dataset and an online platform for reproducible research.

Main Methods:

• A thorough review of existing image deraining approaches was conducted.
• A unified evaluation framework was developed to standardize performance assessment.
• A new high-quality benchmark dataset, HQ-RAIN, comprising 5,000 high-resolution synthetic images, was constructed.
• An online toolkit was developed to facilitate the reproduction and tracking of deraining technologies.

Main Results:

• The study establishes a unified evaluation setting for image deraining methods.
• The HQ-RAIN benchmark provides a high-quality, realistic dataset for extensive evaluations.
• The developed online platform offers an off-the-shelf toolkit for large-scale performance evaluation.

Conclusions:

• A standardized approach to evaluating image deraining methods is crucial for fair comparison and practical application.
• The HQ-RAIN benchmark and online toolkit facilitate reproducible research and track the latest advancements in image deraining.
• Future research should address identified challenges and explore new opportunities in image deraining technology.