Scholarlink

<p>Most existing anomaly detection methods adopt a “one-category-one-model” paradigm, requiring hundreds or even thousands of images per category. This is costly and impractical in industrial settings with limited samples, motivating few-shot anomaly detection (FSAD) to detect anomalies using only a few normal images per category. However, current state-of-the-art FSAD methods fail because direct image comparison overlooks feature-level dependencies, reducing precision. To address these challenges, we propose adaptive graph feature registration (AGFReg). First, we introduce a multi-scale maximum mean discrepancy (MMMD) loss function to measure feature distribution differences across multiple scales, improving robustness while preserving accuracy. Then, we design a graph feature registration module (GFRM) that combines graph convolution and channel attention to effectively capture relationships between features and their neighbors. Experimental results on the MVTec AD and MPDD benchmarks under various few-shot settings demonstrate that AGFReg surpasses existing FSAD methods, yielding an improvement of <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(1.7\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>1.7</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>–<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(10.02\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>10.02</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> in image-level AUC.</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

Few-shot anomaly detection via adaptive graph feature registration

  • 期刊论文

  • 发表日期：

    2026年1月22日

查看全文
  • Gaihua Wang,
  • Sijia Xu,
  • Kehong Li,
  • Keke Ren

摘要

Most existing anomaly detection methods adopt a “one-category-one-model” paradigm, requiring hundreds or even thousands of images per category. This is costly and impractical in industrial settings with limited samples, motivating few-shot anomaly detection (FSAD) to detect anomalies using only a few normal images per category. However, current state-of-the-art FSAD methods fail because direct image comparison overlooks feature-level dependencies, reducing precision. To address these challenges, we propose adaptive graph feature registration (AGFReg). First, we introduce a multi-scale maximum mean discrepancy (MMMD) loss function to measure feature distribution differences across multiple scales, improving robustness while preserving accuracy. Then, we design a graph feature registration module (GFRM) that combines graph convolution and channel attention to effectively capture relationships between features and their neighbors. Experimental results on the MVTec AD and MPDD benchmarks under various few-shot settings demonstrate that AGFReg surpasses existing FSAD methods, yielding an improvement of \(1.7\%\) 1.7 % – \(10.02\%\) 10.02 % in image-level AUC.