Image Deblurring With Image Blurring.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Deep learning (DL) methods show promise for motion deblurring but struggle with complex real-world blurs.
• Existing DL models often perform poorly on real-world data and are sensitive to blur estimation errors, leading to artifacts.

Purpose of the Study:

• To develop a robust motion deblurring framework capable of handling complex real-world blurs.
• To overcome limitations of synthetic datasets and inaccurate blur estimation in current deblurring techniques.

Main Methods:

• Proposed a framework comprising a Blur Space Disentangled Network (BSDNet) and a Hierarchical Scale-recurrent Deblurring Network (HSDNet).
• BSDNet disentangles blur features for improved blur transfer and dataset augmentation.
• HSDNet utilizes disentangled blur features for coarse-to-fine deblurring, breaking the task into subtasks.
• A synthetic motion blur dataset was generated using BSDNet to bridge the gap between training and real-world data.

Main Results:

• The proposed method demonstrated effectiveness on complex real-world blur scenarios.
• BSDNet facilitated better blur feature learning and dataset generation.
• HSDNet successfully recovered sharp details by leveraging blur features and hierarchical processing.
• The framework significantly outperformed existing state-of-the-art deblurring approaches.

Conclusions:

• The developed motion deblurring framework effectively addresses challenges posed by complex real-world blurs and inaccurate blur estimation.
• The combination of BSDNet and HSDNet offers a significant advancement in image deblurring performance.