Scholarlink

<p>Image anomaly detection aims to identify regions deviating from expected visual patterns, with applications spanning industrial defect detection, medical lesion identification, and intelligent security. Traditional methods, focusing on holistic image representation, often struggle with anomalies driven by single latent factors. To address this, we introduce DAD (disentangled anomaly detection), a framework that disentangles object structure from appearance for progressive anomaly detection. DAD employs an abnormal-to-normal transformer (A2NFormer) for structure prediction, followed by an appearance reconstruction module guided by the predicted structure. Finally, an anomaly localization subnetwork combines structural and appearance cues for precise detection. Experimental results on the MVTec Anomaly Detection dataset demonstrate that DAD achieves state-of-the-art performance, improving image-level AUROC by 0.2%, increasing pixel-level AP by 2.5%, and boosting instance-level IAP and IAP@90 by 5.8% and 4.8%, respectively. Our approach highlights the benefits of disentangled representation and structure-guided localization for accurate, robust, and interpretable anomaly detection. The code is publicly available at <a href="https://github.com/yan-yp/DAD">https://github.com/yan-yp/DAD</a> (DOI: <a href="https://doi.org/10.5281/zenodo.15901469">https://doi.org/10.5281/zenodo.15901469</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

Progressive disentanglement for robust image anomaly detection

  • 期刊论文

  • 发表日期：

    2026年2月15日

查看全文
  • Yaping Yan,
  • Hao Chen,
  • Zhenhua Guo,
  • Yuanrui Zeng,
  • Songlin Du

摘要

Image anomaly detection aims to identify regions deviating from expected visual patterns, with applications spanning industrial defect detection, medical lesion identification, and intelligent security. Traditional methods, focusing on holistic image representation, often struggle with anomalies driven by single latent factors. To address this, we introduce DAD (disentangled anomaly detection), a framework that disentangles object structure from appearance for progressive anomaly detection. DAD employs an abnormal-to-normal transformer (A2NFormer) for structure prediction, followed by an appearance reconstruction module guided by the predicted structure. Finally, an anomaly localization subnetwork combines structural and appearance cues for precise detection. Experimental results on the MVTec Anomaly Detection dataset demonstrate that DAD achieves state-of-the-art performance, improving image-level AUROC by 0.2%, increasing pixel-level AP by 2.5%, and boosting instance-level IAP and IAP@90 by 5.8% and 4.8%, respectively. Our approach highlights the benefits of disentangled representation and structure-guided localization for accurate, robust, and interpretable anomaly detection. The code is publicly available at https://github.com/yan-yp/DAD (DOI: https://doi.org/10.5281/zenodo.15901469).