Scholarlink

<p>This study introduces the Normal Label Propagation Algorithm for zero-shot Anomaly Detection (NLPA-AD), a novel approach for anomaly detection in texture-based images without prior training. NLPA-AD mimics human perception by leveraging texture homogeneity to define regions of normality and establish a conceptual understanding of normalcy. The core of NLPA-AD involves a single-label-propagation mechanism that identifies normal areas within the image. This mechanism effectively prevents the propagation of normal labels into anomalous regions by introducing non-propagable points. The identified normal regions are then utilized to construct a normal feature gallery, which serves as a reference for subsequent anomaly detection tasks. To enhance NLPA-AD’s performance, we propose two complementary strategies: Padding and Filtering (P&amp;F) and Dynamic Threshold Matrix Adjustments (DTMA). These strategies address the challenges posed by anomalies that often occur around the periphery of the similarity matrix, effectively mitigating their impact on the overall detection accuracy. Experimental results on the MVTec-AD and DAGM-2007 datasets demonstrate the superiority of NLPA-AD in a zero-shot setting. Specifically, NLPA-AD achieves impressive performance metrics, including 99.7% image-level AUROC and 98.98% image-level F1-max on five textures from the MVTec-AD dataset, and 99.91% image-level AUROC and 98.77% image-level F1-max on nine textures from the DAGM-2007 dataset. These results highlight the effectiveness of NLPA-AD in addressing the challenge of zero-shot anomaly recognition and localization in texture images.</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

NLPA-AD: normal label propagation algorithm for zero-shot texture anomaly detection

  • 期刊论文

  • 发表日期：

    2025年7月10日

查看全文
  • Jiajun Zhang,
  • Yanzhi Song,
  • Zhouwang Yang,
  • Chencheng Wang

摘要

This study introduces the Normal Label Propagation Algorithm for zero-shot Anomaly Detection (NLPA-AD), a novel approach for anomaly detection in texture-based images without prior training. NLPA-AD mimics human perception by leveraging texture homogeneity to define regions of normality and establish a conceptual understanding of normalcy. The core of NLPA-AD involves a single-label-propagation mechanism that identifies normal areas within the image. This mechanism effectively prevents the propagation of normal labels into anomalous regions by introducing non-propagable points. The identified normal regions are then utilized to construct a normal feature gallery, which serves as a reference for subsequent anomaly detection tasks. To enhance NLPA-AD’s performance, we propose two complementary strategies: Padding and Filtering (P&F) and Dynamic Threshold Matrix Adjustments (DTMA). These strategies address the challenges posed by anomalies that often occur around the periphery of the similarity matrix, effectively mitigating their impact on the overall detection accuracy. Experimental results on the MVTec-AD and DAGM-2007 datasets demonstrate the superiority of NLPA-AD in a zero-shot setting. Specifically, NLPA-AD achieves impressive performance metrics, including 99.7% image-level AUROC and 98.98% image-level F1-max on five textures from the MVTec-AD dataset, and 99.91% image-level AUROC and 98.77% image-level F1-max on nine textures from the DAGM-2007 dataset. These results highlight the effectiveness of NLPA-AD in addressing the challenge of zero-shot anomaly recognition and localization in texture images.