Scholarlink

<p>The complexity and openness of railway turnout environments pose great challenges to anomaly detection, and supervised methods are highly dependent on labels, making it difficult to address the diverse types of anomalies and the scarcity of samples in turnout environments. To solve these problems, this paper proposes a new method, Rail-PatchCore, which is based on unsupervised learning and effectively reduces the interference of background noise and enhances the ability to capture anomalous features by adding a Dual-Dimensional Channel Attention (DDCA) module and a projection anomaly scoring module to the PatchCore model. The experiments on our railway-turnout anomaly detection dataset(RTAD) and other datasets (RSDDs, MVTec-AD, BTAD, AEBAD-S) show that the detection performance of Rail-PatchCore is better than that of the existing methods, and the image-level and pixel-level AUCROC indices of Rail-PatchCore on the railway turnout anomaly detection dataset reach 72.2% and 95.3%, respectively. This approach provides an efficient and reliable solution for anomaly detection in railway turnout environments.</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

Rail-PatchCore: unsupervised learning-based detection of visual anomalies in the railway-turnout environment

  • 期刊论文

  • 发表日期：

    2025年2月5日

查看全文
  • YuanHao Zhang,
  • Zujun Yu,
  • Liqiang Zhu,
  • Baoqing Guo,
  • Yao Wang

摘要

The complexity and openness of railway turnout environments pose great challenges to anomaly detection, and supervised methods are highly dependent on labels, making it difficult to address the diverse types of anomalies and the scarcity of samples in turnout environments. To solve these problems, this paper proposes a new method, Rail-PatchCore, which is based on unsupervised learning and effectively reduces the interference of background noise and enhances the ability to capture anomalous features by adding a Dual-Dimensional Channel Attention (DDCA) module and a projection anomaly scoring module to the PatchCore model. The experiments on our railway-turnout anomaly detection dataset(RTAD) and other datasets (RSDDs, MVTec-AD, BTAD, AEBAD-S) show that the detection performance of Rail-PatchCore is better than that of the existing methods, and the image-level and pixel-level AUCROC indices of Rail-PatchCore on the railway turnout anomaly detection dataset reach 72.2% and 95.3%, respectively. This approach provides an efficient and reliable solution for anomaly detection in railway turnout environments.