Scholarlink

<p>In the industrial image anomaly detection area, because of the variety and limited availability of anomalous samples, traditional manual labeling techniques are not only expensive and inefficient, but also not suitable for large-scale production applications. Consequently, reconstruction-based unsupervised learning approaches have gained prominence in this field. However, these methods may not be able to effectively capture all types of anomalies, especially some small anomalies. To address these issues, we propose the Reconstruction Discrimination Network with Adversarial Training (RDNA) for image anomaly detection. RDNA consists of two modules: reconstruction and discriminant modules. Each module is composed of a discriminator and a reconstruction network. Through the adversarial training of the discriminator, the reconstruction module can generate images more consistent with the normal data distribution, and the discriminator module can create more precise anomaly segmentation maps, thereby enhancing the effectiveness of detecting and locating anomalies. Experimental results demonstrate the strong effectiveness in the MVTec AD dataset, achieving an image-level AUROC of 98.5%, a pixel-level AUROC of 97.6%, and an AUPR of 72.8%. Additionally, we conduct experiments on the BTAD dataset, achieving an image-level AUROC of 91.8%. The results also demonstrate that RDNA has strong anomaly detection capabilities, contributing to the field of anomaly detection. Compared to previous methods, RDNA achieves more powerful detection performance. Our implementation is publicly available at <a href="https://github.com/SVIL2024/RDNA">https://github.com/SVIL2024/RDNA</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

RDNA: reconstruction discriminant networks with adversarial training for anomaly detection

  • 期刊论文

  • 发表日期：

    2026年5月5日

查看全文
  • Xiaoru Liu,
  • Shifeng Li,
  • Yan Cheng,
  • Xi Luo

摘要

In the industrial image anomaly detection area, because of the variety and limited availability of anomalous samples, traditional manual labeling techniques are not only expensive and inefficient, but also not suitable for large-scale production applications. Consequently, reconstruction-based unsupervised learning approaches have gained prominence in this field. However, these methods may not be able to effectively capture all types of anomalies, especially some small anomalies. To address these issues, we propose the Reconstruction Discrimination Network with Adversarial Training (RDNA) for image anomaly detection. RDNA consists of two modules: reconstruction and discriminant modules. Each module is composed of a discriminator and a reconstruction network. Through the adversarial training of the discriminator, the reconstruction module can generate images more consistent with the normal data distribution, and the discriminator module can create more precise anomaly segmentation maps, thereby enhancing the effectiveness of detecting and locating anomalies. Experimental results demonstrate the strong effectiveness in the MVTec AD dataset, achieving an image-level AUROC of 98.5%, a pixel-level AUROC of 97.6%, and an AUPR of 72.8%. Additionally, we conduct experiments on the BTAD dataset, achieving an image-level AUROC of 91.8%. The results also demonstrate that RDNA has strong anomaly detection capabilities, contributing to the field of anomaly detection. Compared to previous methods, RDNA achieves more powerful detection performance. Our implementation is publicly available at https://github.com/SVIL2024/RDNA.