FreeDehaze: Towards Training-Free Real-World Image Dehazing via Diffusion Degradation Prior

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Image restoration from hazy conditions is difficult due to complex real-world haze.
• Current methods struggle to accurately model diverse haze representations.

Purpose of the Study:

• To introduce FreeDehaze, a novel training-free diffusion method for real-world image dehazing.
• To leverage text-to-image diffusion models' ability to represent haze effectively.

Main Methods:

• A posterior-based framework that avoids extra degradation estimation networks.
• Generates pseudo-clean images from textual descriptions.
• Uses optimal transport to align denoising output within a PCA-based haze subspace.

Main Results:

• FreeDehaze outperforms comparative methods on subjective metrics across challenging datasets (RTTS, URHI, O-HAZE).
• Achieves competitive objective metrics, demonstrating strong generalization without additional training.
• Effectively addresses non-linear dehazing challenges.

Conclusions:

• FreeDehaze offers a powerful, training-free solution for real-world image dehazing.
• The method's approach, inspired by human cognition, enhances image restoration quality.
• Diffusion models show promise for internalizing and utilizing haze representations in image processing.