From Global to Local: Multi-Patch and Multi-Scale Contrastive Similarity Learning for Unsupervised Defocus Blur...

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Defocus blur detection (DBD) is crucial for various vision tasks.
• Manual pixel-level annotations for training DBD models are labor-intensive and costly.
• Unsupervised DBD methods are gaining importance to overcome annotation limitations.

Purpose of the Study:

• To propose a novel deep network for unsupervised defocus blur detection.
• To eliminate the need for extensive pixel-level manual annotations.
• To enhance the accuracy and efficiency of blur detection in single images.

Main Methods:

• A Multi-patch and Multi-scale Contrastive Similarity (M2CS) learning framework is introduced.
• Composite images are generated by transferring estimated clear/unclear areas.
• Joint global and local similarity discriminators are employed for contrastive learning at multiple scales.

Main Results:

• The proposed M2CS method demonstrates superior performance in unsupervised DBD.
• Experimental results on real-world datasets show significant improvements in quantification and visualization.
• The method effectively distinguishes in-focus and out-of-focus regions without manual labels.

Conclusions:

• The M2CS learning approach provides an effective solution for unsupervised defocus blur detection.
• The joint global-local contrastive strategy enhances the model's ability to handle complex blur scenarios.
• This work contributes a valuable tool for computer vision applications requiring accurate blur assessment.