Scholarlink

<p>Anomaly detection has found extensive applications in industrial manufacturing, medical and other sectors. Traditional unsupervised methods often neglect the issue of global long-range dependency links and local fine-grained information for data. In this study, we introduce the channel attention memory selection network (CAMSN) to efficiently capture hidden features between different pixels in the data. It integrates an attention memory module to enhance the modeling of long-range dependencies in the input data. We propose an adaptive channel selection module to optimize feature transfer and minimize redundant information interference. Comparative experiments are designed on the MVTec AD, CIFAR10 and KolektorSDD datasets. Compared to the latest methods, the results demonstrate an average improvement of 5–12% in AUC scores with 92.32% on the MVTec AD dataset. Ablation studies confirm the efficacy of our method, while analysis of various hyperparameters elucidates impact on experimental outcomes.</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

A channel attention memory selection network for unsupervised anomaly detection

  • 期刊论文

  • 发表日期：

    2025年9月3日

查看全文
  • Faming Gong,
  • Yi Yu,
  • Chengze Du,
  • Junjie Hao,
  • Yingchao Feng,
  • Ya Li

摘要

Anomaly detection has found extensive applications in industrial manufacturing, medical and other sectors. Traditional unsupervised methods often neglect the issue of global long-range dependency links and local fine-grained information for data. In this study, we introduce the channel attention memory selection network (CAMSN) to efficiently capture hidden features between different pixels in the data. It integrates an attention memory module to enhance the modeling of long-range dependencies in the input data. We propose an adaptive channel selection module to optimize feature transfer and minimize redundant information interference. Comparative experiments are designed on the MVTec AD, CIFAR10 and KolektorSDD datasets. Compared to the latest methods, the results demonstrate an average improvement of 5–12% in AUC scores with 92.32% on the MVTec AD dataset. Ablation studies confirm the efficacy of our method, while analysis of various hyperparameters elucidates impact on experimental outcomes.