Research on Surface Defect Detection of Strip Steel Based on Improved YOLOv7

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Surface defect detection is critical for strip steel quality control.
• Existing methods struggle with diverse defect types, scales, and irregular distributions, hindering rapid and accurate detection.
• Challenges include varied defect textures and complex environmental factors.

Purpose of the Study:

• To develop a real-time, high-precision surface defect detection algorithm for strip steel.
• To enhance the efficiency and accuracy of defect identification in industrial settings.
• To address limitations of current methods in detecting diverse and irregularly distributed defects.

Main Methods:

• Utilized YOLOv7 as the baseline architecture.
• Incorporated Partial Convolution (Partial Conv) in the backbone network to reduce model size and increase detection speed.
• Integrated the CA attention mechanism into the ELAN module to improve feature extraction capabilities in complex environments.
• Implemented the SPD convolution module at the output to enhance the detection of small surface defects.

Main Results:

• Achieved a mean average precision (mAP@IoU = 0.5) of 80.4% on the NEU-DET dataset, a 4.0% improvement over the baseline.
• Reduced the number of network parameters by 8.9%.
• Decreased computational load by 21.9% (GFLOPs) while reaching a detection speed of 90.9 FPS.

Conclusions:

• The proposed YOLOv7-based algorithm offers a significant advancement in real-time strip steel surface defect detection.
• The modifications effectively improve detection accuracy, speed, and efficiency, particularly for small and complex defects.
• The algorithm meets the stringent requirements for real-time quality control in strip steel production.