Deep self-supervised spatial-variant image deblurring.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Existing image deblurring methods often rely on synthetic data, limiting real-world performance.
• Real-world blur is frequently spatial-variant and difficult to replicate synthetically.
• The need for blur-sharp pairs hinders the application of current deblurring techniques.

Purpose of the Study:

• To develop a self-supervised learning-based image deblurring method.
• To address limitations of existing methods by handling spatial-variant blur.
• To eliminate the requirement for blur-sharp training pairs.

Main Methods:

• Proposed a novel self-supervised learning framework for image deblurring.
• Introduced a Deblurring Network (D-Net) and a Spatial Degradation Network (SD-Net).
• Utilized an off-the-shelf pre-trained model as a prior and incorporated a recursive optimization strategy.

Main Results:

• The method effectively handles both uniform and spatial-variant blur distributions.
• Achieved favorable performance compared to existing image deblurring methods in extensive experiments.
• Demonstrated the efficacy of the proposed self-supervised approach without synthetic training data.

Conclusions:

• The proposed self-supervised method offers a robust solution for real-world image deblurring.
• The dual-network architecture and optimization strategy contribute to improved deblurring performance.
• This approach advances the field by removing the dependency on synthetic training data.