Scholarlink

<p>With the widespread deployment of machine learning models in real-world scenarios, their transferability to specific environments has become increasingly critical. Take visual anomaly detection (VAD) in robotic overhead line inspection as an example: although existing methods perform well under controlled conditions, their practical effectiveness remains limited due to the high diversity and unpredictability of real-world anomalies. The CableInspect-AD dataset, a real-world cable inspection benchmark, captures actual operational conditions across three cable states: no aging, non-uniform aging, and uniform aging. With potential defects present in all states that may affect safe operation. However, the challenges for anomaly detection vary significantly across these aging conditions. Current general purpose models achieve suboptimal performance, especially on non-uniformly aged cables, and suffer from high false alarm rates, indicating substantial room for improvement. To address this issue, we propose an enhanced framework based on INP-Former, the current state-of-the-art model for anomaly detection. Specifically, we introduce a Multi-Granularity Spatial Aggregation (MGSA) module to extract richer contextual features at different scales. This is followed by a Mutual Voting Mechanism (MVM) that effectively distinguishes normal from anomalous samples by leveraging complementary predictions, thereby significantly reducing false positives caused by complex aging textures. Experimental results show that our method substantially improves anomaly detection performance on non-uniformly aged cables in the CableInspect-AD dataset. It achieves a 2.3% improvement in image-level AUROC and a 1.6% gain in pixel-level AUROC over the original INP-Former, demonstrating its effectiveness and robustness in complex industrial 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

A framework for unsupervised anomaly detection in real-world cable inspection

  • 期刊论文

  • 发表日期：

    2026年1月16日

查看全文
  • Tongwei Lu,
  • Pengyan Shao

摘要

With the widespread deployment of machine learning models in real-world scenarios, their transferability to specific environments has become increasingly critical. Take visual anomaly detection (VAD) in robotic overhead line inspection as an example: although existing methods perform well under controlled conditions, their practical effectiveness remains limited due to the high diversity and unpredictability of real-world anomalies. The CableInspect-AD dataset, a real-world cable inspection benchmark, captures actual operational conditions across three cable states: no aging, non-uniform aging, and uniform aging. With potential defects present in all states that may affect safe operation. However, the challenges for anomaly detection vary significantly across these aging conditions. Current general purpose models achieve suboptimal performance, especially on non-uniformly aged cables, and suffer from high false alarm rates, indicating substantial room for improvement. To address this issue, we propose an enhanced framework based on INP-Former, the current state-of-the-art model for anomaly detection. Specifically, we introduce a Multi-Granularity Spatial Aggregation (MGSA) module to extract richer contextual features at different scales. This is followed by a Mutual Voting Mechanism (MVM) that effectively distinguishes normal from anomalous samples by leveraging complementary predictions, thereby significantly reducing false positives caused by complex aging textures. Experimental results show that our method substantially improves anomaly detection performance on non-uniformly aged cables in the CableInspect-AD dataset. It achieves a 2.3% improvement in image-level AUROC and a 1.6% gain in pixel-level AUROC over the original INP-Former, demonstrating its effectiveness and robustness in complex industrial environments.