MGDR-YOLO: An Efficient Multi-Backbone YOLOv11 Framework for X-Ray Weld Defect Inspection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Materials Science

Background:

• X-ray weld seam imaging presents challenges like weak contrast, slender structures, and multi-scale features.
• Existing detectors struggle with accurate and efficient detection of subtle weld defects.

Purpose of the Study:

• To develop an industrially deployable detector, MGDR-YOLO, for improved X-ray weld defect detection.
• To enhance the accuracy, robustness, and real-time performance of weld defect identification.

Main Methods:

• Proposed MGDR-YOLO with four key innovations: MultiBackbone, Gated Attention Fusion Block (GAFB), Directional Feature Convolution (DFConv), and Rep Shared Convolutional Detection Head (RSCD).
• MultiBackbone enables complementary direction-detail and context modeling.
• GAFB facilitates selective feature fusion, DFConv optimizes directional feature extraction, and RSCD enhances detection head efficiency.

Main Results:

• MGDR-YOLO achieved a mean average precision (mAP) of 95.2%, outperforming YOLOv11n (92.9%).
• Significant mAP improvement of 10.1 percentage points for low-contrast (LP) defects.
• Increased frames per second (FPS) by 39.4% while reducing parameters by 46.2%.

Conclusions:

• MGDR-YOLO offers superior accuracy and robustness in X-ray weld defect detection.
• The model maintains real-time performance, suitable for resource-constrained industrial applications.
• Demonstrates effectiveness in detecting challenging, slender, and low-contrast weld defects.