Scholarlink

<p>In the X-ray digital images of small-diameter copper tube soldering structures, the significant variations in defect size, morphology, and contrast make automatic defect recognition challenging. To address this problem, this study proposes M-KANUNet, an improved TransUNet-based defect segmentation model, designed for segmenting lack of penetration defects in X-ray images of copper tube welds. First, the ResNet module in the encoder part of the baseline TransUNet model is replaced with a Res2Net module combined with the SimAM attention mechanism, enhancing the model’s ability to extract multi-scale features and capture local defect details in complex backgrounds. Second, leveraging the strengths of the Kolmogorov–Arnold Networks (KANs), the KAN–Transformer module is proposed to improve the model’s ability to process complex data by using learnable activation functions at the network’s edges. Experimental results on a constructed dataset of 496 X-ray images show that, compared to the baseline TransUNet model, M-KANUNet achieves improvements of 1.84% in intersection over union (IoU), 1.24% in recall (R), and 2.29% in precision (P), demonstrating higher reliability. Additionally, the frames per second(FPS) of M-KANUNet increases by 10 f/s, and the number of model parameters is reduced by 79.87M. The source code is available at <a href="https://github.com/Lihua-Y/M-KANUNet.">https://github.com/Lihua-Y/M-KANUNet.</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

M-KANUNet: enhanced defect segmentation in X-ray images of copper pipe welds via multi-scale representation and Kolmogorov–Arnold Networks

  • 期刊论文

  • 发表日期：

    2025年2月4日

查看全文
  • Lihua Yang,
  • Jinxian Zhao,
  • Ziming Wang,
  • Yuheng Liu,
  • Dazhao Chi

摘要

In the X-ray digital images of small-diameter copper tube soldering structures, the significant variations in defect size, morphology, and contrast make automatic defect recognition challenging. To address this problem, this study proposes M-KANUNet, an improved TransUNet-based defect segmentation model, designed for segmenting lack of penetration defects in X-ray images of copper tube welds. First, the ResNet module in the encoder part of the baseline TransUNet model is replaced with a Res2Net module combined with the SimAM attention mechanism, enhancing the model’s ability to extract multi-scale features and capture local defect details in complex backgrounds. Second, leveraging the strengths of the Kolmogorov–Arnold Networks (KANs), the KAN–Transformer module is proposed to improve the model’s ability to process complex data by using learnable activation functions at the network’s edges. Experimental results on a constructed dataset of 496 X-ray images show that, compared to the baseline TransUNet model, M-KANUNet achieves improvements of 1.84% in intersection over union (IoU), 1.24% in recall (R), and 2.29% in precision (P), demonstrating higher reliability. Additionally, the frames per second(FPS) of M-KANUNet increases by 10 f/s, and the number of model parameters is reduced by 79.87M. The source code is available at https://github.com/Lihua-Y/M-KANUNet.