Deep Ranking Exemplar-Based Dynamic Scene Deblurring.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Dynamic scene deblurring is mathematically challenging.
• Deep convolutional neural networks (CNNs) have advanced deblurring but often lack structural detail preservation.
• Existing end-to-end CNNs primarily use regression, which can degrade image quality.

Purpose of the Study:

• To propose an exemplar-based method for dynamic scene deblurring.
• To enhance structural detail recovery in deblurred images.
• To overcome limitations of regression-based CNN approaches.

Main Methods:

• Developed a siamese encoder network for input feature extraction.
• Developed a shallow encoder network for exemplar feature extraction.
• Introduced rank modules applied to the last three encoder layers to refine features for deblurring.
• Extended the method for multi-scale processing to recover finer textures.

Main Results:

• The proposed exemplar-based method significantly improves dynamic scene deblurring.
• Achieved notable enhancements in both quantitative and qualitative evaluations.
• Demonstrated superior preservation of structural details compared to existing methods.

Conclusions:

• The exemplar-based approach offers a robust solution for dynamic scene deblurring.
• The integration of siamese and shallow encoders with rank modules effectively enhances blur removal.
• The multi-scale extension allows for richer texture recovery, leading to higher quality deblurred images.