Scholarlink

<p>Nondestructive Evaluation (NDE) is essential for ensuring the reliability of industrial products, but deep learning–based visual inspection systems often fail when applied to new production conditions. These failures arise from domain shifts, such as changes in lighting, sensors, or materials, and correcting them usually requires costly re-annotation of data. To address this challenge, we present an unsupervised domain adaptation (UDA) framework tailored for one-class anomaly detection in Automated Visual Inspection (AVI). Our approach combines adversarial feature alignment with statistical distribution matching, allowing the model to learn domain-invariant representations while maintaining high sensitivity to defects. Evaluation on the MVTec Anomaly Detection benchmark shows that the proposed method substantially reduces performance loss under domain shifts, achieving AUROC above 0.95 and PRO scores above 0.97 across multiple cross-domain scenarios. These findings demonstrate the practical value of our framework and establish a foundation for deploying reliable, scalable, and expert-level anomaly detection solutions in dynamic industrial environments.</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

Unsupervised Domain Adaptation for Visual Anomaly Detection in Nondestructive Evaluation

  • 期刊论文

  • 发表日期：

    2026年4月15日

查看全文
  • Hien Vu Pham,
  • Phan Xuan Tan,
  • Minhhuy Le,
  • Ngoc-Tam Bui

摘要

Nondestructive Evaluation (NDE) is essential for ensuring the reliability of industrial products, but deep learning–based visual inspection systems often fail when applied to new production conditions. These failures arise from domain shifts, such as changes in lighting, sensors, or materials, and correcting them usually requires costly re-annotation of data. To address this challenge, we present an unsupervised domain adaptation (UDA) framework tailored for one-class anomaly detection in Automated Visual Inspection (AVI). Our approach combines adversarial feature alignment with statistical distribution matching, allowing the model to learn domain-invariant representations while maintaining high sensitivity to defects. Evaluation on the MVTec Anomaly Detection benchmark shows that the proposed method substantially reduces performance loss under domain shifts, achieving AUROC above 0.95 and PRO scores above 0.97 across multiple cross-domain scenarios. These findings demonstrate the practical value of our framework and establish a foundation for deploying reliable, scalable, and expert-level anomaly detection solutions in dynamic industrial environments.