A high precision and lightweight method for steel surface defect detection based on improved YOLOv5

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Accurate steel surface defect detection is crucial for safety and efficiency.
• Existing methods face challenges with small flaws, complex surfaces, and real-time processing.

Purpose of the Study:

• To develop an enhanced YOLOv5 model, ASFRW-YOLO, for improved steel surface defect detection.
• To address limitations in accuracy, complexity handling, and real-time performance of current systems.

Main Methods:

• Proposed ASFRW-YOLO model integrating a multi-scale ASF module and RepNCSPELAN4 module.
• Utilized Wise-IoU loss with adaptive weighting for refined bounding box regression.
• Conducted experiments on the NEU-DET dataset using an 8:1:1 train-validation-test split.

Main Results:

• Achieved 83.2% mean Average Precision at IoU 0.5 and 46.4% across IoU 0.5-0.95.
• Demonstrated a ~7 percentage point improvement over YOLOv5s.
• Maintained a lightweight design (6.20M parameters) with ~125 FPS processing for 640x640 images.

Conclusions:

• ASFRW-YOLO effectively balances detection accuracy, computational efficiency, and model compactness.
• The model is highly suitable for real-time industrial defect inspection applications.
• ASFRW-YOLO offers a robust solution for detecting subtle surface flaws in steel.