Advancing Real-World Image Dehazing: Perspective, Modules, and Training.

Area of Science:

• Computer Vision
• Image Processing

Background:

• Deep learning models excel with synthetic data but struggle with real-world hazy images.
• Existing methods often fail to address multiple degradation factors present in real-world hazy images.

Purpose of the Study:

• To develop a robust computational framework for enhancing image quality from degraded hazy observations.
• To bridge the domain gap between synthetic and real-world image dehazing.
• To improve the adaptability and performance of deep models in real-world scenarios.

Main Methods:

• Developed a novel hazy imaging model to simulate multiple degradation attributes.
• Designed a "localization-and-removal" dehazing network with degradation localization and removal modules.
• Introduced a Gaussian perceptual contrastive loss for natural dehazing direction.

Main Results:

• The proposed method demonstrates superior performance on challenging real-world hazy datasets (RTTS, URHI, Fattal).
• Outperforms current state-of-the-art methods in both objective image quality metrics and subjective visual effects.
• Effectively avoids spurious correlations in feature extraction through a novel degradation removal module.

Conclusions:

• The new framework significantly enhances real-world image dehazing capabilities.
• The "localization-and-removal" approach combined with a new loss function offers a promising direction for future research.
• The developed hazy imaging model aids in creating more realistic training data.