Scholarlink

<p>In recent years, unsupervised anomaly detection has made significant advancements. However, its performance is often contingent on mild assumptions, demonstrating limited robustness to variations in sample scale and position. Moreover, it faces considerable challenges when domain shifts occur between normal samples in the training and test sets. These limitations hinder its practical applicability in complex industrial environments. To address these challenges, we propose a method that combines masked perceiving distillation and enhanced pyramid feature reconstruction, leveraging the modeling of relatively invariant conditional distributions within image patches to mitigate the impact of domain shifts. Specifically, we improve the model’s ability to perceive samples of varying sizes and positions through dense feature extraction across different receptive fields. Additionally, we utilize wavelet-enhanced multi-scale contextual information to optimize the reconstruction process, thereby further reducing sensitivity to domain shifts. On the most challenging AeBAD-S dataset, our method outperforms existing state-of-the-art approaches, achieving a detection AUROC score of 88.0% and a localization PRO score of 90.8%, with improvements of 3.3% and 1.7%, respectively. Moreover, our model achieves superior performance across multiple categories of the MVTec AD dataset (including its corrupted variants), and the MPDD dataset, further demonstrating its strong robustness and generalization capability.</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

RMR: Robust industrial anomaly detection based on the masked perceiving distillation and enhanced feature reconstruction

  • 期刊论文

  • 发表日期：

    2025年5月23日

查看全文
  • Nan Zhou,
  • Shaofei Wu

摘要

In recent years, unsupervised anomaly detection has made significant advancements. However, its performance is often contingent on mild assumptions, demonstrating limited robustness to variations in sample scale and position. Moreover, it faces considerable challenges when domain shifts occur between normal samples in the training and test sets. These limitations hinder its practical applicability in complex industrial environments. To address these challenges, we propose a method that combines masked perceiving distillation and enhanced pyramid feature reconstruction, leveraging the modeling of relatively invariant conditional distributions within image patches to mitigate the impact of domain shifts. Specifically, we improve the model’s ability to perceive samples of varying sizes and positions through dense feature extraction across different receptive fields. Additionally, we utilize wavelet-enhanced multi-scale contextual information to optimize the reconstruction process, thereby further reducing sensitivity to domain shifts. On the most challenging AeBAD-S dataset, our method outperforms existing state-of-the-art approaches, achieving a detection AUROC score of 88.0% and a localization PRO score of 90.8%, with improvements of 3.3% and 1.7%, respectively. Moreover, our model achieves superior performance across multiple categories of the MVTec AD dataset (including its corrupted variants), and the MPDD dataset, further demonstrating its strong robustness and generalization capability.