An improved steel defect detection model using multi-scale feature fusion based on YOLO-MFD.

Area of Science:

• Computer Vision
• Materials Science
• Artificial Intelligence

Background:

• Steel surface defect detection faces challenges with small targets and complex backgrounds.
• Existing models often have large parameter counts, hindering deployment.
• Accurate and efficient detection is crucial for industrial quality control.

Purpose of the Study:

• To develop a lightweight and accurate steel surface defect detection model.
• To address limitations in detecting small defects and reduce model complexity.
• To enable real-time defect detection in industrial settings.

Main Methods:

• Proposed YOLO-MFD model based on YOLO11n.
• Introduced Multi-scale Edge Feature Enhancement (MEFE) module for subtle defect detection.
• Implemented Global-to-Local Spatial Attention Bidirectional Feature Pyramid Network (GLSA-BiFPN) for efficient feature fusion.
• Utilized a Group detection head (Ghead) to reduce parameters via grouped convolution.

Main Results:

• YOLO-MFD achieved an mAP of 83.8% on the NEU-DET dataset, a 5.9% improvement over YOLO11n.
• Reduced model parameters by 27% and computational cost by 6.3%.
• Demonstrated effective balance between detection accuracy and lightweight design.

Conclusions:

• The proposed YOLO-MFD model significantly enhances steel surface defect detection accuracy and efficiency.
• The model's lightweight nature makes it suitable for real-time deployment in industrial scenarios.
• YOLO-MFD offers a practical solution for complex industrial quality inspection.