Scholarlink

<p>In recent years, anomaly detection techniques have found extensive utilization across diverse domains and various industrial applications. Traditional anomaly detection methods using distillation models, employ a teacher-student framework where the teacher model extracts features and the student model learns to reconstruct normal patterns for anomaly detection. However, distillation methods still face significant challenges in complex scenarios: the scarcity of anomalous samples and high annotation costs lead to insufficient model generalization; the diversity and complexity of anomaly types make it difficult for traditional methods to distinguish between different anomalies; and the blurred boundaries between normal and anomalous samples hinder the detection of subtle anomalous patterns. To address these challenges, this paper introduces an auxiliary classifier in the conventional distillation framework with reverse data flow. Using intermediate features of the student model, the proposed method captures fine-grained differences between various types of anomalies and improves the discriminative ability between normal and anomalous samples through supervised learning, thus improving both the precision of anomaly detection. Moreover, this reverse distillation flow effectively identifies the normal data distribution by extracting multilevel features from pre-trained models and reconstructing input data, enabling more accurate anomaly detection. Compared to traditional distillation approaches, our method demonstrates superior feature extraction capabilities and better adaptability to scenarios with limited annotated data, offering an effective solution for anomaly detection in complex environments. Our source code and data are available online at <a href="https://github.com/wkwj888/AD.git.">https://github.com/wkwj888/AD.git.</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

Reverse distillation with auxiliary contrastive learning for robust industrial anomaly detection

  • 期刊论文

  • 发表日期：

    2025年12月22日

查看全文
  • Huabo Shen,
  • Hua Yang,
  • Xiaodong Sun,
  • Kai Wang

摘要

In recent years, anomaly detection techniques have found extensive utilization across diverse domains and various industrial applications. Traditional anomaly detection methods using distillation models, employ a teacher-student framework where the teacher model extracts features and the student model learns to reconstruct normal patterns for anomaly detection. However, distillation methods still face significant challenges in complex scenarios: the scarcity of anomalous samples and high annotation costs lead to insufficient model generalization; the diversity and complexity of anomaly types make it difficult for traditional methods to distinguish between different anomalies; and the blurred boundaries between normal and anomalous samples hinder the detection of subtle anomalous patterns. To address these challenges, this paper introduces an auxiliary classifier in the conventional distillation framework with reverse data flow. Using intermediate features of the student model, the proposed method captures fine-grained differences between various types of anomalies and improves the discriminative ability between normal and anomalous samples through supervised learning, thus improving both the precision of anomaly detection. Moreover, this reverse distillation flow effectively identifies the normal data distribution by extracting multilevel features from pre-trained models and reconstructing input data, enabling more accurate anomaly detection. Compared to traditional distillation approaches, our method demonstrates superior feature extraction capabilities and better adaptability to scenarios with limited annotated data, offering an effective solution for anomaly detection in complex environments. Our source code and data are available online at https://github.com/wkwj888/AD.git.