YOLO-DCF: dual distillation and context-aware fusion for defect detection

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Industrial surface defect detection is crucial for manufacturing quality across sectors like steel and electronics.
• Challenges include diverse defect types, scale variations, and background noise, hindering practical deployment.

Purpose of the Study:

• To develop a novel, lightweight detection framework, YOLO-DCF, to address challenges in industrial surface defect detection.
• To enhance detection precision, robustness, and real-time inference capabilities for manufacturing quality inspection.

Main Methods:

• Proposed YOLO-DCF framework built upon YOLO11, incorporating Context-Guided Dynamic Fusion FPN, C3k2-Dilated Multiscale Contextual Residual module, and Dual Block-Channel Knowledge Distillation.
• Context-Guided Dynamic Fusion FPN decomposes global context for precise defect localization and noise suppression.
• Dual Block-Channel Knowledge Distillation enhances model compression via self-distillation, preserving essential representations.

Main Results:

• YOLO-DCF achieved mAP50 scores of 79.3% on NEU-DET and 96.5% on PKU-Market-PCB, outperforming baseline methods.
• Demonstrated improved recall and robustness for fine-grained and low-contrast defects.
• Maintained competitive real-time inference capability despite increased model complexity.

Conclusions:

• YOLO-DCF offers a practical and deployable solution for industrial quality inspection.
• The framework provides efficient, distribution-aware visual recognition for manufacturing contexts.
• This work advances the field of automated visual inspection in industrial settings.