Scholarlink

<p>Defects inspection is critical to guarantee the product quality in industrial production, especially for metal products such as rail steel and magnetic tile. Although existing deep learning methods have achieved many successes, their performance will suffer great degradation caused by weak edge features, tiny defect, intraclass diversity and interclass resemblance on metal surface defects. In addition, most methods fail to meet the real-time detection need in real scenarios. In this study, a novel detection method for metal products based on Transformer and edge attention mechanism (EAformer) is proposed. Firstly, a low-level feature enhancement module (LFEM) is proposed to enlarge the receptive field (RF) and strengthen the contextual information of shallow features. Secondly, an edge enhancement attention module (EEAM) is proposed to enhance semantic representation by fusing boundary features of defect part and residual area from different hierarchies. Finally, an asymmetric convolutional feature fusion module (ACFFM) is proposed to extract effective feature information of defects with variable size. The mIoU can be increased by 1.72%, 1.32%, and 1.51% by EAformer on three typical metal defect datasets, respectively, compared with the suboptimal method (i.e., SegDINO). Moreover, EAformer achieves 30.26 FPS, which indicates that it can satisfy the demand of real-time defect detection.</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

Edge attention-based transformer for metal surface defect segmentation

  • 期刊论文

  • 发表日期：

    2026年2月1日

查看全文
  • Lijun Kong,
  • Jie Duan,
  • Jia Chen,
  • Yulong Zhang,
  • Lixiang Zhao,
  • Jianbo Yu

摘要

Defects inspection is critical to guarantee the product quality in industrial production, especially for metal products such as rail steel and magnetic tile. Although existing deep learning methods have achieved many successes, their performance will suffer great degradation caused by weak edge features, tiny defect, intraclass diversity and interclass resemblance on metal surface defects. In addition, most methods fail to meet the real-time detection need in real scenarios. In this study, a novel detection method for metal products based on Transformer and edge attention mechanism (EAformer) is proposed. Firstly, a low-level feature enhancement module (LFEM) is proposed to enlarge the receptive field (RF) and strengthen the contextual information of shallow features. Secondly, an edge enhancement attention module (EEAM) is proposed to enhance semantic representation by fusing boundary features of defect part and residual area from different hierarchies. Finally, an asymmetric convolutional feature fusion module (ACFFM) is proposed to extract effective feature information of defects with variable size. The mIoU can be increased by 1.72%, 1.32%, and 1.51% by EAformer on three typical metal defect datasets, respectively, compared with the suboptimal method (i.e., SegDINO). Moreover, EAformer achieves 30.26 FPS, which indicates that it can satisfy the demand of real-time defect detection.