Scholarlink

<p>In response to the challenges associated with radiographic weld inspection, such as multi-scale defects, image quality heterogeneity, and background interference, this study introduces an advanced weld defect detection framework. This framework integrates the YOLOv11 detector with SAHI slicing. A weld region extraction strategy employing watershed segmentation and frequency-domain filtering is utilized to suppress irrelevant background elements. Concurrently, an adaptive enhancement pipeline, which combines median filtering, non-local means denoising, and CLAHE, is implemented to improve image quality. The SAHI framework, featuring dynamic slicing, facilitates robust multi-scale defect detection, while a two-stage label fusion process reconstructs fragmented defects and resolves cross-category ambiguities. The method was evaluated on an original dataset encompassing five defect categories, achieving a precision of 94.9%, recall of 86.4%, mAP@0.5 of 95.1%, and mAP@[0.5:0.95] of 71.4%, thereby demonstrating high detection capability. When tested on an independent external dataset, the method maintained robust performance, with a precision of 78.8%, recall of 72.8%, mAP@0.5 of 77.8%, and mAP@[0.5:0.95] of 46.9%, confirming its effectiveness and generalizability in practical applications.</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

SAHI-Enhanced YOLOv11 with Dual-Stage Label Fusion for Intelligent Weld Defect Detection in X-Ray Radiographs

  • 期刊论文

  • 发表日期：

    2026年6月16日

查看全文
  • Jiaying Yan,
  • Gefei Kuang,
  • Chenggang Wang,
  • Rui Feng,
  • Linhao Wang,
  • Jiahao Zhang,
  • Mingjun Zhang,
  • Tao Liu

摘要

In response to the challenges associated with radiographic weld inspection, such as multi-scale defects, image quality heterogeneity, and background interference, this study introduces an advanced weld defect detection framework. This framework integrates the YOLOv11 detector with SAHI slicing. A weld region extraction strategy employing watershed segmentation and frequency-domain filtering is utilized to suppress irrelevant background elements. Concurrently, an adaptive enhancement pipeline, which combines median filtering, non-local means denoising, and CLAHE, is implemented to improve image quality. The SAHI framework, featuring dynamic slicing, facilitates robust multi-scale defect detection, while a two-stage label fusion process reconstructs fragmented defects and resolves cross-category ambiguities. The method was evaluated on an original dataset encompassing five defect categories, achieving a precision of 94.9%, recall of 86.4%, mAP@0.5 of 95.1%, and mAP@[0.5:0.95] of 71.4%, thereby demonstrating high detection capability. When tested on an independent external dataset, the method maintained robust performance, with a precision of 78.8%, recall of 72.8%, mAP@0.5 of 77.8%, and mAP@[0.5:0.95] of 46.9%, confirming its effectiveness and generalizability in practical applications.