Scholarlink

<p>Deep learning techniques have revolutionized industrial anomaly detection. However, their reliance on substantial labeled data poses challenges in scenarios where anomalous samples are scarce. This paper presents a practical self-supervised anomaly detection framework, CutMask-based Anomaly Detection (CMAD), designed for industrial inspection using normal target-domain training images together with independently sourced defect-shape priors. The main methodological contribution is CutMask, a defect-shape-aware data augmentation strategy that, unlike generic augmentations such as CutMix, MixUp, and regular CutPaste-style rectangular patch pasting, explicitly leverages obtainable prior knowledge of defect morphology to generate more realistic and morphology-consistent simulated defect samples. Furthermore, CMAD uses a lightweight ResNet-18 as the backbone for anomaly-sensitive feature extraction, integrates a Self-Supervised Predictive Convolutional Attentive Block (SSPCAB) module to refine feature modeling, and optimizes the network with the proposed Cyclical Auxiliary Focal Loss (CAFL) objective to improve discrimination between normal and CutMask-generated pseudo-anomalous samples. Experimental results on the MVTec AD dataset, the VisA dataset, and a practical presswork dataset show that CMAD achieves strong image-level anomaly detection performance under limited anomalous data conditions, including image-level area under the receiver operating characteristic curve (AUROC) scores of 97.93% on MVTec AD and 93.88% on VisA. Overall, the proposed framework provides a practical and defect-shape-aware extension of augmentation-driven industrial anomaly detection for scenarios in which real anomalous samples are scarce.</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

Enhanced industrial anomaly detection via CutMask data augmentation: a self-supervised approach

  • 期刊论文

  • 发表日期：

    2026年6月23日

查看全文
  • Le Yang,
  • Dingjian Yao,
  • Wenhan Yang,
  • Zhengsong Wang

摘要

Deep learning techniques have revolutionized industrial anomaly detection. However, their reliance on substantial labeled data poses challenges in scenarios where anomalous samples are scarce. This paper presents a practical self-supervised anomaly detection framework, CutMask-based Anomaly Detection (CMAD), designed for industrial inspection using normal target-domain training images together with independently sourced defect-shape priors. The main methodological contribution is CutMask, a defect-shape-aware data augmentation strategy that, unlike generic augmentations such as CutMix, MixUp, and regular CutPaste-style rectangular patch pasting, explicitly leverages obtainable prior knowledge of defect morphology to generate more realistic and morphology-consistent simulated defect samples. Furthermore, CMAD uses a lightweight ResNet-18 as the backbone for anomaly-sensitive feature extraction, integrates a Self-Supervised Predictive Convolutional Attentive Block (SSPCAB) module to refine feature modeling, and optimizes the network with the proposed Cyclical Auxiliary Focal Loss (CAFL) objective to improve discrimination between normal and CutMask-generated pseudo-anomalous samples. Experimental results on the MVTec AD dataset, the VisA dataset, and a practical presswork dataset show that CMAD achieves strong image-level anomaly detection performance under limited anomalous data conditions, including image-level area under the receiver operating characteristic curve (AUROC) scores of 97.93% on MVTec AD and 93.88% on VisA. Overall, the proposed framework provides a practical and defect-shape-aware extension of augmentation-driven industrial anomaly detection for scenarios in which real anomalous samples are scarce.