Scholarlink

<p>Weld proximity defects often occur in the steel-building sheet’s welding process, presenting characteristics such as denseness, area occupancy, extension, and similarity, which are challenging for accurate detection in industrial welding systems. Existing defect detectors are focused on small-sized and individual targets, which may not be suitable for detecting weld proximity defects. Due to the mentioned issues, we propose the SEP-YOLOv7 model. The improved Efficient Channel Attention (ECA) mechanism based on the Average Sigmoid-Tanh kernel bias and the one-dimensional Conv_BN_SiLU are introduced into the Enhanced Local Attention Network of YOLOv7, and then the additional SiLU activation is added at the last layer of the ECA structure, aimed to increase the feature extraction and differentiation ability. Next, the multi-filter convolution residual block and partially decoupled convolution layer are designed in the prediction Head to recognize and classify weld proximity defects accurately. We adopt the steel-building sheet welding dataset and the NEU-DET dataset to demonstrate the model’s effect and generalization. By comparing the related detectors with the proposed SEP-YOLOv7, experimental results show that our proposed model presents superior effects, with mAP 86.4/75.8 pct, Precision 87.9/74.7 pct, and Average 82.8/70.2 pct.</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

Average Sigmoid-Tanh Attention and Multi-filter Partially Decoupled Mechanism via YOLOv7 for Detecting Weld Proximity Defects

  • 期刊论文

  • 发表日期：

    2025年6月16日

查看全文
  • Zihua Chen,
  • Runmei Zhang,
  • Meng-Yen Hsieh,
  • Alireza Souri,
  • Kuan-Ching Li

摘要

Weld proximity defects often occur in the steel-building sheet’s welding process, presenting characteristics such as denseness, area occupancy, extension, and similarity, which are challenging for accurate detection in industrial welding systems. Existing defect detectors are focused on small-sized and individual targets, which may not be suitable for detecting weld proximity defects. Due to the mentioned issues, we propose the SEP-YOLOv7 model. The improved Efficient Channel Attention (ECA) mechanism based on the Average Sigmoid-Tanh kernel bias and the one-dimensional Conv_BN_SiLU are introduced into the Enhanced Local Attention Network of YOLOv7, and then the additional SiLU activation is added at the last layer of the ECA structure, aimed to increase the feature extraction and differentiation ability. Next, the multi-filter convolution residual block and partially decoupled convolution layer are designed in the prediction Head to recognize and classify weld proximity defects accurately. We adopt the steel-building sheet welding dataset and the NEU-DET dataset to demonstrate the model’s effect and generalization. By comparing the related detectors with the proposed SEP-YOLOv7, experimental results show that our proposed model presents superior effects, with mAP 86.4/75.8 pct, Precision 87.9/74.7 pct, and Average 82.8/70.2 pct.