Related Experiment Video
Updated: Jan 13, 2026
07:58
Automatic Laser-based Geometry Capture for Finite Element Analysis of Weld Beads
Published on: July 25, 2025
751
DEIM-SFA: A Multi-Module Enhanced Model for Accurate Detection of Weld Surface Defects
Yan Sun1, Yingjie Xie1, Ran Peng1
1College of Information Engineering, Sichuan Agricultural University, Ya'an 625000, China.
Sensors (Basel, Switzerland)
|October 29, 2025
Summary
This study introduces DEIM-SFA, a new framework for automated welding defect detection. It significantly improves accuracy by better retaining fine-grained features and fusing multi-scale information, enhancing structural safety in manufacturing.
Area of Science:
- Manufacturing
- Materials Science
- Computer Vision
Background:
- Automated detection of metal welding defects is crucial for manufacturing.
- Existing methods lack fine-grained feature retention and efficient multi-scale fusion, leading to low accuracy.
- Complex backgrounds often interfere with defect detection systems.
Purpose of the Study:
- To develop a novel detection framework, DEIM-SFA, for high-precision automated visual inspection of welding defects.
- To address limitations in fine-grained feature retention, multi-scale information fusion, and background interference.
Main Methods:
- Introduced structure-aware dynamic convolution (SPD-Conv) to focus on defect structures and suppress noise.
- Designed a multi-scale dynamic fusion pyramid (FTPN) for efficient aggregation of multi-scale features.
- Integrated a lightweight multi-scale attention module (EMA) to enhance salient region localization.
Main Results:
- DEIM-SFA achieved significant improvements: 3.9% in mAP50, 4.3% in mAP75, 3.7% in mAP50-95, and 1.4% in Recall.
- Demonstrated superior detection accuracy across various target sizes.
- Maintained balanced model complexity and efficient inference compared to SOTA methods.
Conclusions:
- The DEIM-SFA framework offers a robust solution for automated welding defect detection.
- The proposed methods effectively enhance feature extraction and fusion for improved accuracy.
- DEIM-SFA surpasses existing methods in detecting welding defects in industrial machine vision.
