Scholarlink

<p>In the field of nondestructive testing, especially in X-ray weld defect detection where computational resources and storage space are limited, traditional target detection models face great challenges in dealing with low-contrast, multi-scale, and morphologically complex weld defects. In this paper, a Lightweight Multi-Scale Real-Time non-convolution detection TRansformer (LMS-RTDETR) is proposed with the aim of improving the detection of weld defects in resource-limited environments. First, a Multi-Scale Feature Aggregation (MSFA) module is used for parallel convolution and feature reorganization. Secondly, an Efficient Additive Attention Feature Interaction module reduces computational complexity from quadratic to linear while maintaining contextual awareness. Then, a Multi-Scale Feature Pyramid Network (MSFPN) implements multi-branch pathways for effective feature fusion. Finally, bounding box regression is optimized using the NWD-Inner GIoU loss function. The experiments show that LMS-RTDETR reduces the model parameters and floating point operations on the X-ray weld defect dataset by 38.34% and 29.30%. mAP50:95 improves by 3.10% points. This study provides a high-precision solution for industrial non-destructive testing.&#xa0;</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

Lightweight DETR algorithm for X-ray weld defect detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月25日

查看全文
  • Xipeng Zhang,
  • Heng Chen,
  • Ziyang Liu,
  • Shang Anli

摘要

In the field of nondestructive testing, especially in X-ray weld defect detection where computational resources and storage space are limited, traditional target detection models face great challenges in dealing with low-contrast, multi-scale, and morphologically complex weld defects. In this paper, a Lightweight Multi-Scale Real-Time non-convolution detection TRansformer (LMS-RTDETR) is proposed with the aim of improving the detection of weld defects in resource-limited environments. First, a Multi-Scale Feature Aggregation (MSFA) module is used for parallel convolution and feature reorganization. Secondly, an Efficient Additive Attention Feature Interaction module reduces computational complexity from quadratic to linear while maintaining contextual awareness. Then, a Multi-Scale Feature Pyramid Network (MSFPN) implements multi-branch pathways for effective feature fusion. Finally, bounding box regression is optimized using the NWD-Inner GIoU loss function. The experiments show that LMS-RTDETR reduces the model parameters and floating point operations on the X-ray weld defect dataset by 38.34% and 29.30%. mAP50:95 improves by 3.10% points. This study provides a high-precision solution for industrial non-destructive testing.