Multi-Type Weld Defect Detection in Galvanized Sheet MIG Welding Using an Improved YOLOv10 Model

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation
• Quality Control

Background:

• Automated weld inspection is challenging due to reflective surfaces, changing textures, and subtle defects in galvanized-sheet MIG welding.
• Existing methods struggle with practical shop-floor constraints, such as limited computational resources (edge devices) and variable lighting conditions.
• The need for robust, lightweight, and accurate weld inspection systems is critical for maintaining high-quality manufacturing.

Purpose of the Study:

• To develop a compact and efficient deep learning model (YOLO-MIG) for accurate weld-seam inspection under realistic production conditions.
• To address the limitations of current vision systems in detecting subtle defects on reflective galvanized steel.
• To enable edge deployment of intelligent weld quality control systems.

Main Methods:

• A novel compact object detection model, YOLO-MIG, was developed based on the YOLOv10n architecture.
• Key modifications include a C2f-EMSCP backbone for preserving weak defect cues, a BiFPN neck for small-target feature fusion, and a C2fCIB head to reduce false positives from seam edges and illumination.
• The model was trained and evaluated on a custom dataset of 2608 augmented images collected from a workshop environment.

Main Results:

• YOLO-MIG achieved high performance with 98.4% mAP@0.5 and 56.29% mAP@0.5:0.95 on the test set.
• The model is lightweight, featuring 1.83 million parameters and 3.87 MB FP16 weights, suitable for edge deployment.
• Compared to the baseline YOLOv10n, YOLO-MIG significantly improved accuracy while reducing model size, parameters, and computational load.

Conclusions:

• YOLO-MIG demonstrates superior accuracy and efficiency for weld-seam inspection in challenging industrial environments.
• The proposed architectural modifications effectively handle reflective surfaces, varying textures, and subtle defects.
• The model's lightweight nature makes it a practical solution for real-time, edge-based intelligent weld quality control.