Stimulating Diffusion Model for Image Denoising via Adaptive Embedding and Ensembling.

Area of Science:

• Computational Photography
• Artificial Intelligence
• Image Processing

Background:

• Image denoising is crucial in computational photography, balancing perceptual quality and distortion.
• Existing methods often compromise either visual fidelity or introduce artifacts.
• Diffusion models show promise but face challenges in direct application to image denoising.

Purpose of the Study:

• To develop a novel strategy for image denoising using diffusion models.
• To address input and content inconsistencies in applying diffusion models to denoising.
• To achieve state-of-the-art performance in both perceptual quality and distortion reduction.

Main Methods:

• Introduced a Diffusion Model for Image Denoising (DMID) strategy.
• Developed an adaptive embedding method for noisy images into pre-trained diffusion models.
• Implemented an adaptive ensembling method to minimize distortion in denoised images.

Main Results:

• DMID achieved state-of-the-art performance on benchmark datasets.
• The strategy demonstrated superior results for both Gaussian and real-world image noise.
• Both distortion-based and perception-based metrics showed significant improvements.

Conclusions:

• The proposed DMID strategy effectively overcomes limitations of current image denoising techniques.
• This approach offers a robust solution for high-quality image denoising using diffusion models.
• DMID represents a significant advancement in computational photography and image restoration.