Scholarlink

<p>Anomaly detection in industrial printing images is challenged by the scarcity of anomalous samples, limiting the application of model-based techniques. To address this, we introduce Un-AMNet, an unsupervised network incorporating an adaptive mechanism for detecting defects in printed labels. Un-AMNet comprises three key components: a shallow feature extractor, an anomaly feature generator and an anomaly feature discriminator. The feature extractor transforms features from the pre-trained backbone network into local features, while the anomaly feature generator constructs synthetic anomalies by introducing noise into the normal feature space, effectively mitigating the problem of sample scarcity. The domain-adaptive discriminator, optimized for barcode characteristics, employs adaptive mechanisms to enhance defect sensitivity. We evaluate Un-AMNet on a custom label defect dataset (CodeDataset) and a public benchmark (MVTec AD), achieving image-level anomaly detection accuracy (I-AUROC) of 95.0<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> and 99.5<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> and pixel-level localization accuracy (P-AUROC) of 94.2<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> and 97.4<InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation>, respectively. These results demonstrate Un-AMNet’s effectiveness and generalizability in anomaly detection tasks. Our code is available at <a href="https://github.com/DL-Mao/Un-AMNet">https://github.com/DL-Mao/Un-AMNet</a>.</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

Adaptive mechanism-based unsupervised network for anomaly detection in printed labels

  • 期刊论文

  • 发表日期：

    2026年3月13日

查看全文
  • Dianlu Hu,
  • Sen Wang,
  • Lun Zhao,
  • Yu Ren,
  • Xue Jing,
  • Xuangang Li,
  • Lan Zhang

摘要

Anomaly detection in industrial printing images is challenged by the scarcity of anomalous samples, limiting the application of model-based techniques. To address this, we introduce Un-AMNet, an unsupervised network incorporating an adaptive mechanism for detecting defects in printed labels. Un-AMNet comprises three key components: a shallow feature extractor, an anomaly feature generator and an anomaly feature discriminator. The feature extractor transforms features from the pre-trained backbone network into local features, while the anomaly feature generator constructs synthetic anomalies by introducing noise into the normal feature space, effectively mitigating the problem of sample scarcity. The domain-adaptive discriminator, optimized for barcode characteristics, employs adaptive mechanisms to enhance defect sensitivity. We evaluate Un-AMNet on a custom label defect dataset (CodeDataset) and a public benchmark (MVTec AD), achieving image-level anomaly detection accuracy (I-AUROC) of 95.0 \(\%\) % and 99.5 \(\%\) % and pixel-level localization accuracy (P-AUROC) of 94.2 \(\%\) % and 97.4 \(\%\) % , respectively. These results demonstrate Un-AMNet’s effectiveness and generalizability in anomaly detection tasks. Our code is available at https://github.com/DL-Mao/Un-AMNet.