Scholarlink

<p>Industrial anomaly detection aims to identify surface defects (e.g., in texture and color) and structural defects in products. A significant obstacle in this field is the necessity for large, high-quality training datasets. While zero-shot and few-shot learning methods have emerged to address this need by operating with minimal labeled data, they are often plagued by slow inference speeds that hinder real-time industrial deployment. To bridge this gap between data efficiency and speed, we propose a novel zero-shot approach that integrates initial data screening with diffusion-based data augmentation. Our framework begins with a similarity-based screening algorithm to curate a set of normal samples directly from the overall data distribution. A latent diffusion model then learns the patch-level distribution of these samples to generate a robustly augmented dataset, which is used to construct a comprehensive memory bank for anomaly detection. Evaluated on the MVTec AD dataset, our method achieves state-of-the-art performance, with average AUROC scores of 96.87% (pixel-level) and 96.25% (image-level). It also exhibits remarkable inter-class stability, as indicated by low inter-class AUROC variances of 5.80 (pixel-level) and 28.53 (image-level). Most critically for industrial applications, our method achieves a detection speed of 220 ms per image, which is approximately 8 times faster than the leading zero-shot method. These results demonstrate the method’s strong generalization capability and its potential for deployment in complex industrial anomaly detection scenarios.</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

Diffusion-Augmented Zero-Shot Industrial Anomaly Detection with Similarity-Based Data Screening

  • 期刊论文

  • 发表日期：

    2026年7月4日

查看全文
  • Yuxin Li,
  • Ruijia Guan,
  • Songtao Ni,
  • Xu Zhao

摘要

Industrial anomaly detection aims to identify surface defects (e.g., in texture and color) and structural defects in products. A significant obstacle in this field is the necessity for large, high-quality training datasets. While zero-shot and few-shot learning methods have emerged to address this need by operating with minimal labeled data, they are often plagued by slow inference speeds that hinder real-time industrial deployment. To bridge this gap between data efficiency and speed, we propose a novel zero-shot approach that integrates initial data screening with diffusion-based data augmentation. Our framework begins with a similarity-based screening algorithm to curate a set of normal samples directly from the overall data distribution. A latent diffusion model then learns the patch-level distribution of these samples to generate a robustly augmented dataset, which is used to construct a comprehensive memory bank for anomaly detection. Evaluated on the MVTec AD dataset, our method achieves state-of-the-art performance, with average AUROC scores of 96.87% (pixel-level) and 96.25% (image-level). It also exhibits remarkable inter-class stability, as indicated by low inter-class AUROC variances of 5.80 (pixel-level) and 28.53 (image-level). Most critically for industrial applications, our method achieves a detection speed of 220 ms per image, which is approximately 8 times faster than the leading zero-shot method. These results demonstrate the method’s strong generalization capability and its potential for deployment in complex industrial anomaly detection scenarios.