Scholarlink

<p>Traditional industrial anomaly detection methods often struggle with high computational complexity, hindering their deployment in resource-constrained environments. To address this limitation, we propose a Target-Oriented Anomaly Synthesis Strategy (TOASS), a lightweight self-supervised anomaly detection framework tailored for industrial applications. TOASS introduces two key innovations: an advanced Feature Fusion Adapter (FFA) that enhances feature representation by transferring local features to the target domain, mitigating domain bias; and an Anomaly Detection Module (ADM) that synthesizes abnormal samples in the feature space by adding Gaussian noise to normal features, enabling self-supervised learning without labeled data. Extensive experiments on the MVTec-AD benchmark demonstrate that TOASS maintains a lightweight design suitable for real-time industrial anomaly detection while achieving performance close to state-of-the-art levels, with a high frame rate of 90 FPS on a 3080Ti GPU and an I-AUROC of 99.5%. The code and dataset are available at: <a href="https://github.com/mostimali/TOASS">https://github.com/mostimali/TOASS</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

Lightweight self-supervised anomaly detection via feature space synthesis for industrial applications

  • 期刊论文

  • 发表日期：

    2025年7月1日

查看全文
  • Shen-Bin Li,
  • Rui-Sheng Jia

摘要

Traditional industrial anomaly detection methods often struggle with high computational complexity, hindering their deployment in resource-constrained environments. To address this limitation, we propose a Target-Oriented Anomaly Synthesis Strategy (TOASS), a lightweight self-supervised anomaly detection framework tailored for industrial applications. TOASS introduces two key innovations: an advanced Feature Fusion Adapter (FFA) that enhances feature representation by transferring local features to the target domain, mitigating domain bias; and an Anomaly Detection Module (ADM) that synthesizes abnormal samples in the feature space by adding Gaussian noise to normal features, enabling self-supervised learning without labeled data. Extensive experiments on the MVTec-AD benchmark demonstrate that TOASS maintains a lightweight design suitable for real-time industrial anomaly detection while achieving performance close to state-of-the-art levels, with a high frame rate of 90 FPS on a 3080Ti GPU and an I-AUROC of 99.5%. The code and dataset are available at: https://github.com/mostimali/TOASS.