Scholarlink

<p>In this paper, we introduce SWRD, a new public dataset containing over 3600 seam weld X-ray images, categorized into standard seam welds and T-joint seam welds. Each image is annotated with polygonal labels for specific defects, making the dataset suitable for various deep learning tasks such as classification, object detection, and instance segmentation. We also detail the defect formation mechanisms and their corresponding characteristics in X-ray images. To enhance the usability of the dataset for deep learning models, we applied several image processing techniques, including image adjustment, sliding window cropping, and preprocessing. Our experiments with the state-of-the-art YOLOv8 object detection models show promising results, with the YOLOv8m model achieving a mAP50 of 0.66 and a mAP50-95 of 0.49. Given that we used default training parameters and limited training epochs, we anticipate even better performance with further optimization. The complete dataset can be downloaded from: <a href="http://www.tz-ndt.com/#/download">http://www.tz-ndt.com/#/download</a>.</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

SWRD: A Dataset of Radiographic Image of Seam Weld for Defect Detection

  • 期刊论文

  • 发表日期：

    2025年5月7日

查看全文
  • Xuefeng Zhao,
  • Juntao Wu,
  • Baoxin Zhang,
  • Haoyu Wen,
  • Xiaopeng Wang,
  • Yan Li,
  • Xinghua Yu

摘要

In this paper, we introduce SWRD, a new public dataset containing over 3600 seam weld X-ray images, categorized into standard seam welds and T-joint seam welds. Each image is annotated with polygonal labels for specific defects, making the dataset suitable for various deep learning tasks such as classification, object detection, and instance segmentation. We also detail the defect formation mechanisms and their corresponding characteristics in X-ray images. To enhance the usability of the dataset for deep learning models, we applied several image processing techniques, including image adjustment, sliding window cropping, and preprocessing. Our experiments with the state-of-the-art YOLOv8 object detection models show promising results, with the YOLOv8m model achieving a mAP50 of 0.66 and a mAP50-95 of 0.49. Given that we used default training parameters and limited training epochs, we anticipate even better performance with further optimization. The complete dataset can be downloaded from: http://www.tz-ndt.com/#/download.