Diffusion Models in Low-Level Vision: A Survey

Area of Science:

• Computer Vision
• Deep Learning
• Generative Models

Background:

• Deep generative models excel in low-level vision.
• Diffusion models are prominent for high-quality image generation.
• A comprehensive survey of diffusion models in low-level vision is lacking.

Purpose of the Study:

• To provide the first comprehensive review of denoising diffusion models in low-level vision.
• To cover theoretical and practical contributions.
• To synthesize advances and identify future research directions.

Main Methods:

• Outlined three general diffusion modeling frameworks.
• Explored connections with other deep generative models.
• Categorized diffusion models by framework and application.
• Reviewed benchmarks and evaluation metrics.
• Evaluated diffusion models across six representative tasks.

Main Results:

• Diffusion models demonstrate powerful generative capabilities in low-level vision.
• Applications span natural image processing, medical imaging, remote sensing, and video processing.
• Extensive evaluation shows quantitative and qualitative performance.

Conclusions:

• Denoising diffusion models are crucial for low-level vision tasks.
• Current limitations exist, with promising future research directions identified.
• This review fosters a deeper understanding of diffusion models in the field.