Scholarlink

<p>In industrial production, the detection of structural anomalies on product surfaces is crucial for quality control. Traditional encoder–decoder architectures for unsupervised anomaly detection, while effective, tend to reduce the reconstruction error between normal and abnormal samples due to their generalization capabilities. To address this, we propose an enhanced structural anomaly detection algorithm based on improved image inpainting. By converting the reconstruction task into an inpainting-filling-reconstruction process, we amplify the reconstruction error between normal and abnormal samples. Our algorithm uses a feature loss and feature-level anomaly discrimination method to mitigate noise interference. Furthermore, we introduce a lightweight U-Net design to meet industrial requirements. Experimental results on the MVTec LOCO AD and MVTec AD datasets demonstrate the effectiveness of our approach. On MVTec LOCO AD, ESAD achieves an average image‑level AUROC of 85.0%, outperforming existing methods on structural anomalies. On the more general MVTec AD benchmark, it attains 95.6% image‑level AUROC and 96.6% pixel‑level AUROC, surpassing several state‑of‑the‑art algorithms. These results highlight the strong generalization and practical potential of ESAD for industrial anomaly 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

Enhanced structural anomaly detection through improved image inpainting and feature-level discrimination

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月6日

查看全文
  • Dongce Fei,
  • Qiliang Wu

摘要

In industrial production, the detection of structural anomalies on product surfaces is crucial for quality control. Traditional encoder–decoder architectures for unsupervised anomaly detection, while effective, tend to reduce the reconstruction error between normal and abnormal samples due to their generalization capabilities. To address this, we propose an enhanced structural anomaly detection algorithm based on improved image inpainting. By converting the reconstruction task into an inpainting-filling-reconstruction process, we amplify the reconstruction error between normal and abnormal samples. Our algorithm uses a feature loss and feature-level anomaly discrimination method to mitigate noise interference. Furthermore, we introduce a lightweight U-Net design to meet industrial requirements. Experimental results on the MVTec LOCO AD and MVTec AD datasets demonstrate the effectiveness of our approach. On MVTec LOCO AD, ESAD achieves an average image‑level AUROC of 85.0%, outperforming existing methods on structural anomalies. On the more general MVTec AD benchmark, it attains 95.6% image‑level AUROC and 96.6% pixel‑level AUROC, surpassing several state‑of‑the‑art algorithms. These results highlight the strong generalization and practical potential of ESAD for industrial anomaly detection.