Scholarlink

<p>In the field of industrial anomaly detection, early anomaly detection algorithms typically assume that all training data remain static during the training process, meaning that samples from all classes or tasks are provided simultaneously and trained jointly. This assumption is idealized, as industrial anomaly detection often involves dynamic changes in the detection environment and the introduction of new class data. Such changes can lead to catastrophic forgetting and performance degradation of the model. To address these challenges, this paper combines class-incremental learning with zero-shot anomaly detection, proposing a novel zero-shot class-incremental anomaly detection method (ZSCIAD). We partition the data from different categories into distinct tasks and train the model in a class-incremental manner. A CNN-based feature extractor is employed to extract features from unlabeled training samples. To tackle the complex anomalies present in certain data, this paper proposes a feature enhancement module based on a self-attention mechanism that utilizes global information to aid in the detection of these anomalies. Furthermore, the paper employs the Local Outlier Factor (LOF) algorithm to detect and eliminate anomalous features in the feature space, followed by feature sampling. We construct a global memory bank to store the sampled feature sets from multiple classes, facilitating effective anomaly detection. Our method demonstrates strong performance on the challenging MVTec AD, BTAD, and DTD-Synthetic datasets. Specifically, it achieves an average classification AUROC of 98% on MVTec AD, 94% on BTAD, and 95% on DTD-Synthetic. Moreover, our approach surpasses several classical methods in terms of stability across different tasks, consistently maintaining high detection accuracy. Code is available at <a href="https://github.com/giraffeGYH/ZSCIAD">https://github.com/giraffeGYH/ZSCIAD</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

ZSCIAD: a preliminary exploration of zero-shot class incremental anomaly detection method

  • 期刊论文

  • 发表日期：

    2025年10月23日

查看全文
  • Yaohua Guo,
  • Guoai Xu,
  • Jianping Yin

摘要

In the field of industrial anomaly detection, early anomaly detection algorithms typically assume that all training data remain static during the training process, meaning that samples from all classes or tasks are provided simultaneously and trained jointly. This assumption is idealized, as industrial anomaly detection often involves dynamic changes in the detection environment and the introduction of new class data. Such changes can lead to catastrophic forgetting and performance degradation of the model. To address these challenges, this paper combines class-incremental learning with zero-shot anomaly detection, proposing a novel zero-shot class-incremental anomaly detection method (ZSCIAD). We partition the data from different categories into distinct tasks and train the model in a class-incremental manner. A CNN-based feature extractor is employed to extract features from unlabeled training samples. To tackle the complex anomalies present in certain data, this paper proposes a feature enhancement module based on a self-attention mechanism that utilizes global information to aid in the detection of these anomalies. Furthermore, the paper employs the Local Outlier Factor (LOF) algorithm to detect and eliminate anomalous features in the feature space, followed by feature sampling. We construct a global memory bank to store the sampled feature sets from multiple classes, facilitating effective anomaly detection. Our method demonstrates strong performance on the challenging MVTec AD, BTAD, and DTD-Synthetic datasets. Specifically, it achieves an average classification AUROC of 98% on MVTec AD, 94% on BTAD, and 95% on DTD-Synthetic. Moreover, our approach surpasses several classical methods in terms of stability across different tasks, consistently maintaining high detection accuracy. Code is available at https://github.com/giraffeGYH/ZSCIAD.