An Unsupervised Image Dehazing With Scene Geometry Prior for Road Traffic Scenarios.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Single image dehazing is challenging for real-world road traffic scenes due to limited paired data and complex scene geometry.
• Existing methods struggle with traffic-specific geometric features and real-time processing constraints.

Purpose of the Study:

• To develop a robust and efficient unsupervised dehazing method for real-world road traffic scenes.
• To introduce a novel scene geometry prior (SGP) that leverages depth cues from vanishing points (VP) for geometry-aware guidance.
• To reduce the reliance on large paired datasets for training dehazing models.

Main Methods:

• Proposed a scene geometry prior (SGP) with global (G-SGP) and non-local (NL-SGP) components to capture and refine geometric distributions.
• Developed a lightweight, unsupervised road traffic image dehazing network (RTDnet) utilizing G-SGP, NL-SGP, and VP information.
• Introduced an atmospheric scattering model (ASM)-driven mutual-boost learning mechanism (ASM-ML) for effective prior integration and physical knowledge distillation.

Main Results:

• RTDnet achieved state-of-the-art performance in restoration quality, efficiency, and model size compared to existing methods.
• The SGP and ASM-ML enabled training without paired traffic data by exploiting traffic-specific geometry.
• The lightweight design facilitates real-time deployment and robust dehazing in challenging conditions.

Conclusions:

• The proposed SGP and RTDnet offer a significant advancement in unsupervised road traffic image dehazing.
• The method effectively addresses limitations of scarce data and complex geometry, enabling real-time applications.
• RTDnet's robust performance enhances the utility of downstream computer vision tasks in hazy environments.