Scholarlink

<p>Thermal data combined with artificial intelligence (AI) has proven to be a promising technology for welding image recognition (WIR). Among deep learning (DL) models, You Only Look Once (YOLO) stands out due to its fast recognition capability, moderate to low computational cost, and simplicity compared to other methods. However, its application has been largely limited to object detection. This work aims to evaluate the prospects of YOLO applied to thermal weld pool images for segmentation and classification tasks. Thermal images were captured in four distinct welding positions, under the same parameter settings, from the initiation of welding until a point within the steady-state (SS) phase. The thermal field was monitored from the backside of fully penetrated welds produced via the TIG process. A subset of SS images, selected based on thermal field distribution characteristics observed in the experiments, was used for model training. Two YOLO nanosized models were investigated to optimize processing speed and resource efficiency. The results highlighted the thermal field’s significance as a valuable data source and demonstrated AI’s potential to extract meaningful insights. YOLO effectively segmented the weld pool in thermal images, even when trained on a small dataset limited to steady-state conditions. For classification tasks, YOLOv8 outperformed YOLO11 in recognizing thermal patterns in the welding experiments. The nano model series demonstrated efficient and accurate predictions with low processing times.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep learning-based YOLO for semantic segmentation and classification of weld pool thermal images

  • 期刊论文

  • 发表日期：

    2025年3月15日

查看全文
  • Vinicius Lemes Jorge,
  • Issam Bendaoud,
  • Fabien Soulié,
  • Cyril Bordreuil

摘要

Thermal data combined with artificial intelligence (AI) has proven to be a promising technology for welding image recognition (WIR). Among deep learning (DL) models, You Only Look Once (YOLO) stands out due to its fast recognition capability, moderate to low computational cost, and simplicity compared to other methods. However, its application has been largely limited to object detection. This work aims to evaluate the prospects of YOLO applied to thermal weld pool images for segmentation and classification tasks. Thermal images were captured in four distinct welding positions, under the same parameter settings, from the initiation of welding until a point within the steady-state (SS) phase. The thermal field was monitored from the backside of fully penetrated welds produced via the TIG process. A subset of SS images, selected based on thermal field distribution characteristics observed in the experiments, was used for model training. Two YOLO nanosized models were investigated to optimize processing speed and resource efficiency. The results highlighted the thermal field’s significance as a valuable data source and demonstrated AI’s potential to extract meaningful insights. YOLO effectively segmented the weld pool in thermal images, even when trained on a small dataset limited to steady-state conditions. For classification tasks, YOLOv8 outperformed YOLO11 in recognizing thermal patterns in the welding experiments. The nano model series demonstrated efficient and accurate predictions with low processing times.