An improved lightweight YOLOv11 algorithm for weld surface defect detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Materials Science

Background:

• Industrial welding faces challenges with unclear defect characteristics and complex backgrounds.
• Existing defect detection models are often costly and lack efficiency in weld surface defect detection.

Purpose of the Study:

• To propose an improved, lightweight YOLOv11 model (YOLO-Air) for enhanced weld surface defect detection.
• To address limitations of existing models regarding cost, complexity, and detection accuracy.

Main Methods:

• Integration of feature extraction and convolutional modules for improved feature representation and efficiency.
• Incorporation of GSConv and VOV-GSCSP modules in the neck network to reduce feature redundancy and computational load.
• Design of a lightweight detection head to further decrease model complexity.

Main Results:

• The YOLO-Air model demonstrated superior performance on the Welding Defect Test-V2 and NEU-DET datasets.
• Achieved a 1.3% improvement in mAP50 metric.
• Reduced the number of parameters by 17.3% and computational complexity by 31.7%.

Conclusions:

• The proposed YOLO-Air model offers an effective and efficient solution for weld surface defect detection.
• The model's lightweight design and improved performance make it a viable option for industrial applications.
• Experimental data confirmed the robustness of the YOLO-Air model's performance.