Swin-Transformer -YOLOv5 for lightweight hot-rolled steel strips surface defect detection algorithm

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Automatic visual inspection of hot-rolled steel strips is crucial for industrial quality control.
• Existing defect detection algorithms face challenges with accuracy, speed, and suitability for low-power platforms.

Purpose of the Study:

• To develop a lightweight and accurate surface defect detection algorithm for hot-rolled steel strips.
• To improve the real-time performance and efficiency of defect identification systems.

Main Methods:

• Proposed an improved Swin-Transformer-YOLOv5 model, a one-stage object detection framework.
• Incorporated GhostNet for model lightweighting, Swin-Transformer within the C3 module to handle complex backgrounds and categories, CoordAttention for enhanced feature extraction, and BiFPN for multi-scale feature fusion.

Main Results:

• The improved model significantly outperforms standard target detection algorithms.
• Achieved an 8.39% increase in mAP (mean Average Precision) compared to the original model.
• Reduced model parameters by 36.6%, GFLOPs by 40.0%, and weight by 34.7%.

Conclusions:

• The developed Swin-Transformer-YOLOv5 model offers superior performance in hot-rolled steel strip surface defect detection.
• The model's efficiency and reduced computational requirements make it ideal for deployment on low-arithmetic platforms.