PDFormer: a polarization-driven dual-branch transformer for dense fog removal

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Fog significantly degrades image quality by reducing contrast and obscuring details due to light scattering.
• Current defogging methods often fail in dense fog due to limitations in physical understanding and global context processing.

Purpose of the Study:

• To develop an advanced image defogging technique that overcomes the limitations of existing methods.
• To improve the restoration of fine-grained details and global context in foggy images.

Main Methods:

• Proposed a novel dual-branch transformer architecture (PDFormer) integrating physics-based and data-driven insights.
• An intensity branch captures long-range dependencies using a hierarchical multiscale window-based Transformer.
• A polarization branch extracts scattering-aware features from multi-angle polarization images, guided by physical priors.

Main Results:

• PDFormer achieved state-of-the-art performance on synthetic and real-world foggy datasets.
• Demonstrated superior perceptual quality and quantitative metrics compared to existing methods.
• Effectively restored visibility across various fog densities.

Conclusions:

• The proposed PDFormer effectively integrates physical priors and data-driven learning for robust image defogging.
• The dual-branch architecture successfully captures both global context and local scattering-aware features.
• PDFormer offers a significant advancement in restoring clear images from challenging foggy scenes.