Defocus Blur Detection Attack via Mutual-Referenced Feature Transfer.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep learning has significantly advanced defocus blur detection (DBD).
• Current DBD methods often rely on multiscale and multilevel features.
• Adversarial attacks offer a new perspective for analyzing and improving DBD networks.

Purpose of the Study:

• To propose a novel mutual-referenced attack framework for defocus blur detection (DBD) networks.
• To generate high-quality adversarial images by utilizing mutual feature references between image regions.
• To explore applications in enhancing DBD performance and defocus deblurring.

Main Methods:

• A divide-and-conquer perturbation image generation model is designed.
• Separate attack images are generated for focus regions and defocus areas.
• Mutual-referenced feature transfer (MRFT) models are integrated to boost attack effectiveness.

Main Results:

• The proposed mutual-referenced attack framework demonstrates effectiveness in generating adversarial images.
• Comprehensive experiments validate the performance of the developed method.
• The study showcases potential applications in sample augmentation for DBD and paired sample generation for deblurring.

Conclusions:

• The mutual-referenced attack framework provides a novel approach to attack DBD networks.
• Leveraging mutual feature references between focus and defocus regions improves adversarial image quality.
• The method has promising applications for improving DBD and defocus deblurring tasks.