Strip Surface Defect Detection Algorithm Based on YOLOv5.

Area of Science:

• Materials Science
• Computer Science
• Artificial Intelligence

Background:

• Surface defect detection in industrial hot rolled strip steel is crucial for quality control.
• Existing methods may lack accuracy and efficiency.
• Deep learning offers advanced solutions for complex detection tasks.

Purpose of the Study:

• To improve the detection accuracy of surface defects in industrial hot rolled strip steel.
• To develop an efficient and accurate deep learning framework for this task.
• To introduce a novel multi-scale feature fusion module for enhanced feature extraction.

Main Methods:

• A novel convolutional neural network (CNN) framework, CG-Net, was developed.
• A multi-scale feature fusion module (ATPF) was proposed to integrate and weight features adaptively.
• The model was trained and tested on strip surface defect datasets, including NEU-CLS.

Main Results:

• CG-Net achieved an average accuracy of 75.9% (mAP50) with 105 frames per second (FPS) and 6.5 GFLOPs.
• Detection accuracy improved by 6.3% compared to the baseline YOLOv5s.
• The model demonstrated reduced reference quantity (67%) and computational load (59.5%) versus YOLOv5s.
• Good generalization performance was observed on the NEU-CLS dataset.

Conclusions:

• The proposed CG-Net framework effectively improves surface defect detection accuracy for industrial hot rolled strip steel.
• The ATPF module enhances the extraction of multi-scale semantic information, leading to better performance.
• CG-Net offers a computationally efficient and accurate solution with strong generalization capabilities.