Frequency compensated diffusion model for real-scene dehazing.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep learning models for image dehazing degrade on real-world data due to distribution shift.
• Standard diffusion models face challenges in learning high-frequency components crucial for detail reconstruction.

Purpose of the Study:

• To develop a robust image dehazing framework with improved generalization to real-world haze.
• To address the spectral bias issue in diffusion models for better image detail recovery.

Main Methods:

• A novel Frequency Compensation block (FCB) was designed to emphasize mid-to-high frequencies.
• A data augmentation pipeline, HazeAug, was introduced to increase haze diversity and degree.
• A conditional diffusion model framework was established for blind dehazing.

Main Results:

• Diffusion models integrated with FCB showed significant improvements in perceptual and distortion metrics.
• The HazeAug pipeline enhanced the generalization capability of the dehazing model.
• The proposed model achieved over 1 dB PSNR improvement on challenging real-world datasets like Dense-Haze and Nh-Haze.

Conclusions:

• The proposed conditional diffusion model with FCB and HazeAug offers superior performance for real-world image dehazing.
• The framework demonstrates strong generalization capabilities, outperforming recent methods.
• This work provides an effective solution for the distribution shift problem in image dehazing.