Scholarlink

<p>Pipeline steel generally exhibits more subtle defect characteristics after surface modification and microstructural modification treatments, making conventional inspection methods insufficient for quality evaluation requirements. A dedicated image dataset containing six typical defect categories was therefore established for petrochemical heat-treated and modified pipeline steel surfaces. Defect types cover crazing, inclusion, patches, pitted surface, rolled-in scale and scratches in actual industrial environments. Six classical models including ResNet-50 V2, MobileNetV3-small, EfficientNetV2-small, ConvNeXt-small, Swin-small and MobileViT-S are compared on established dataset. Comprehensive experimental results verify that Swin-small performs best among all selected baseline models in classification tasks. Novel improved model named CC-Swin-small is designed on original Swin-small framework for pipeline steel defect classification. Cross-shaped window Attention and Convolutional Token Embedding are embedded into backbone structure to strengthen feature representation ability. Defect classification accuracy was significantly improved by the proposed model, while inference latency remained within an acceptable range for industrial applications. Final experimental results demonstrate that CC-Swin-small achieves an overall classification accuracy of 97.59% on constructed dataset, which significantly exceeds 88.52% yielded by original Swin-small model under identical experimental settings.</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

Enhanced Swin Transformer for High-Accuracy Detection and Classification of Pipeline Steel Surface Defects

  • 期刊论文

  • 发表日期：

    2026年7月24日

查看全文
  • Jiayi Li,
  • Wencan Xiong,
  • Junming Lu,
  • Ping Gu,
  • Yong He,
  • Ji Chen,
  • Xiaoxiao Dong

摘要

Pipeline steel generally exhibits more subtle defect characteristics after surface modification and microstructural modification treatments, making conventional inspection methods insufficient for quality evaluation requirements. A dedicated image dataset containing six typical defect categories was therefore established for petrochemical heat-treated and modified pipeline steel surfaces. Defect types cover crazing, inclusion, patches, pitted surface, rolled-in scale and scratches in actual industrial environments. Six classical models including ResNet-50 V2, MobileNetV3-small, EfficientNetV2-small, ConvNeXt-small, Swin-small and MobileViT-S are compared on established dataset. Comprehensive experimental results verify that Swin-small performs best among all selected baseline models in classification tasks. Novel improved model named CC-Swin-small is designed on original Swin-small framework for pipeline steel defect classification. Cross-shaped window Attention and Convolutional Token Embedding are embedded into backbone structure to strengthen feature representation ability. Defect classification accuracy was significantly improved by the proposed model, while inference latency remained within an acceptable range for industrial applications. Final experimental results demonstrate that CC-Swin-small achieves an overall classification accuracy of 97.59% on constructed dataset, which significantly exceeds 88.52% yielded by original Swin-small model under identical experimental settings.