Detection method of small size defects on pipeline weld surface based on improved YOLOv7

Area of Science:

• Computer Vision
• Machine Learning
• Materials Science

Background:

• Pipeline weld surface defect images present complex backgrounds and small defect sizes.
• Small defects are prone to missed detection and false positives in automated inspection.

Purpose of the Study:

• To develop a lightweight target detection algorithm for accurately identifying small defects in pipeline weld images.
• To improve the efficiency and reliability of automated weld inspection systems.

Main Methods:

• An improved YOLOv7 algorithm incorporating a 160*160 small target detection head was developed.
• Depthwise separable convolutions replaced standard convolutions to reduce computational overhead.
• The CIoU loss function was optimized to the EIoU loss function for faster model convergence.

Main Results:

• The improved YOLOv7 algorithm achieved a mean average precision (mAP@0.5) of 72.2%, an 11% increase over the original YOLOv7.
• Model computation and parameter count were reduced by 75.6% and 60.3%, respectively.
• The algorithm demonstrated high confidence in detecting small-sized defects.

Conclusions:

• The proposed lightweight YOLOv7 algorithm effectively detects small defects on pipeline weld surfaces.
• The enhancements lead to superior detection performance and reduced computational resources.
• This method offers a promising solution for automated inspection in industrial settings.