Steel surface defect detection based on multi-layer fusion networks.

Area of Science:

• Materials Science and Engineering
• Computer Vision and Machine Learning
• Industrial Quality Control

Background:

• Accurate steel surface defect detection is vital for industrial quality and safety.
• Traditional methods struggle with complex defect shapes and low-resolution images.
• Deep learning, specifically YOLOv5, shows promise but requires enhancement for challenging scenarios.

Purpose of the Study:

• To improve the precision and recall of steel surface defect detection using deep learning.
• To enhance the YOLOv5 model's capability in identifying defects with complex shapes and small sizes.
• To develop a more efficient and effective object detection algorithm for industrial applications.

Main Methods:

• Integration of the RepBi-PAN fusion network into the YOLOv5 architecture.
• Optimization of the model backbone using DenseNet for superior feature extraction.
• Incorporation of the Normalized Attention Module (NAM) to boost small target detection.

Main Results:

• Achieved a 4.1% increase in mean average precision (mAP), 3.2% in precision, and 2.4% in recall.
• Demonstrated superior performance over the original YOLOv5 in complex backgrounds and small target recognition.
• Outperformed other YOLO algorithms in recall and mAP with a smaller model size.

Conclusions:

• The enhanced YOLOv5 model significantly improves steel surface defect detection accuracy and efficiency.
• The proposed method offers a competitive alternative to existing YOLO algorithms, including YOLOv9, with fewer parameters and lower computational cost.
• This advancement contributes to more reliable quality control and safety in steel manufacturing.