Reliable image dehazing by NeRF.

Area of Science:

• Computer Vision
• Image Processing
• Computer Graphics

Background:

• Image dehazing is crucial for low-level vision tasks.
• Existing methods struggle with real-world datasets, unreliable processes, and complex lighting.
• There's a need for robust dehazing solutions and high-quality datasets.

Purpose of the Study:

• To propose a new haze model integrating optical scattering and computer graphics rendering.
• To develop a versatile dataset generation pipeline for image dehazing.
• To enhance dehazing algorithms with improved visual effects and objective indicators.

Main Methods:

• A novel haze model combining optical scattering and computer graphics rendering was developed.
• A dataset generation pipeline using 3D fog space reconstruction and voxel deletion was created.
• Unreal Engine 5 and laboratory experiments were used for simulation and validation.

Main Results:

• The proposed pipeline generates high-quality dehazed images and datasets without paired training or prior knowledge.
• Effective dehazing was achieved across diverse and complex outdoor lighting conditions.
• A voxel-based data enhancement method was introduced.

Conclusions:

• The developed haze model and generation pipeline offer a reliable solution for image dehazing.
• The approach is suitable for state-of-the-art dehazing algorithms, improving visual and objective performance.
• This work provides a valuable resource for advancing image dehazing research.