Real-world defocus deblurring via score-based diffusion models.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Defocus blur, a common artifact from camera depth-of-field limitations, poses challenges for image restoration.
• Existing deep learning methods for defocus deblurring often require paired all-in-focus and defocus images, which are difficult to obtain in real-world scenarios.

Purpose of the Study:

• To propose a high-resolution iterative deblurring method for real scenes using a score-based diffusion model.
• To overcome the limitations of paired training data dependency in traditional defocus deblurring techniques.

Main Methods:

• Training a score network to learn the score function of focused images at various noise levels.
• Utilizing a reverse-time stochastic differential equation (SDE) for image reconstruction.
• Implementing a prediction-correction (PC) framework to enhance robustness by correcting SDE discretization errors.
• Employing an iterative approach inherent to diffusion models for gradual image quality enhancement.

Main Results:

• The proposed method achieves superior performance on real-world datasets, including self-captured data, RealBlur, and DED.
• Demonstrated improvements of approximately 13.4% in Peak Signal-to-Noise Ratio (PSNR) and 34.7% in Structural Similarity Index Measure (SSIM) compared to DnCNN, FFDNet, and CycleGAN.
• Successfully enabled high-resolution iterative defocus deblurring without paired training data.

Conclusions:

• The score-based diffusion model offers an effective solution for high-resolution iterative defocus deblurring in real-world scenarios.
• The method significantly enhances image clarity and robustness, addressing the limitations of current approaches.
• This work paves the way for more accessible and accurate image restoration techniques in practical applications.