Strip steel surface defect detection based on lightweight YOLOv5.

Area of Science:

• Materials Science
• Computer Science
• Artificial Intelligence

Background:

• Deep learning methods for strip steel surface defect detection face challenges like target loss, false alarms, and computational demands.
• Existing algorithms struggle to balance detection accuracy with real-time processing speed.

Purpose of the Study:

• To develop a lightweight YOLOv5 algorithm for strip steel surface defect detection that optimizes the trade-off between accuracy and speed.
• To address limitations of current deep learning models in industrial defect detection.

Main Methods:

• Introduced GSConv, an efficient lightweight convolutional layer, and a Slim Neck architecture to reduce parameters and enhance speed.
• Integrated SimAM, a non-parametric attention mechanism, into the neck to improve detection accuracy.
• Employed the SIoU loss function for regression prediction to accelerate convergence and boost efficiency.

Main Results:

• The proposed YOLOv5-GSS algorithm achieved 83.8% mean average precision (mAP) on the NEU-DET dataset.
• The model demonstrated a detection speed of 100 Hz, which is 3.8 Hz faster than the standard YOLOv5.
• Outperformed the baseline YOLOv5 method by 2.9% in accuracy.

Conclusions:

• The YOLOv5-GSS algorithm effectively balances detection accuracy and speed for strip steel surface defect detection.
• The optimization strategy, incorporating GSConv, SimAM, and SIoU loss, proves efficacious and superior to existing methods.
• The developed model offers a practical and efficient solution for industrial surface defect inspection.