Scholarlink

<p>This study presents a&#xa0;neural network model for automatic defect detection and localization in terahertz images of manufactured components. The model employs an architecture based on MBConv blocks and achieves mean IoU values of 0.93 and 0.72 for two defect types, respectively. To eliminate the class imbalance, a&#xa0;15-fold cross-validation with a&#xa0;weighted loss of the cross entropy is used. Training was performed using the AdamW optimizer with OneCycleLR scheduling and BF16 precision. The results obtained demonstrate that the model can accurately segment defects of various types and sizes, making it suitable for practical non-destructive testing 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

Development and training of a neural network model for semantic segmentation in terahertz non-destructive testing

  • 期刊论文

  • 发表日期：

    2025年11月1日

查看全文
  • D. A. Shatokhina,
  • A. K. Berzin,
  • A. V. Badin,
  • A. I. Berdyugin,
  • K. V. Dorozhkin,
  • M. S. Yuzhakov,
  • E. A. Trofimov,
  • A. V. Perevalov

摘要

This study presents a neural network model for automatic defect detection and localization in terahertz images of manufactured components. The model employs an architecture based on MBConv blocks and achieves mean IoU values of 0.93 and 0.72 for two defect types, respectively. To eliminate the class imbalance, a 15-fold cross-validation with a weighted loss of the cross entropy is used. Training was performed using the AdamW optimizer with OneCycleLR scheduling and BF16 precision. The results obtained demonstrate that the model can accurately segment defects of various types and sizes, making it suitable for practical non-destructive testing applications.