Scholarlink

In light of the current challenges associated with low detection accuracy for anomalous samples and the significant difficulty in collecting these samples, this study employs an unsupervised learning approach to enhance both detection accuracy and speed. Specifically, the SC-ResNet feature extraction module is utilized for anomaly detection to extract channel information and enrich feature representation. Additionally, an improved S-Coordinate Attention mechanism is integrated into the model, which emphasizes spatial information, allowing the model to focus on significant features. The model employs a novel method for identifying abnormal data. This paper presents a comparative analysis, focusing on the performance metrics of the new model against several state-of-the-art techniques. Experimental results indicate that the proposed model demonstrates exceptional performance in terms of accuracy and reliability. Within the MVTEC dataset, seven categories achieve 100% accuracy, while the average AUROC for object categories reaches 99.6%. This indicates that the proposed model exhibits strong robustness in anomaly detection across diverse object categories.

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

Anomaly Detection in Industrial Images Based on Unsupervised Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Hao Wang,
  • Jianjun Li,
  • Yong Wang

摘要

In light of the current challenges associated with low detection accuracy for anomalous samples and the significant difficulty in collecting these samples, this study employs an unsupervised learning approach to enhance both detection accuracy and speed. Specifically, the SC-ResNet feature extraction module is utilized for anomaly detection to extract channel information and enrich feature representation. Additionally, an improved S-Coordinate Attention mechanism is integrated into the model, which emphasizes spatial information, allowing the model to focus on significant features. The model employs a novel method for identifying abnormal data. This paper presents a comparative analysis, focusing on the performance metrics of the new model against several state-of-the-art techniques. Experimental results indicate that the proposed model demonstrates exceptional performance in terms of accuracy and reliability. Within the MVTEC dataset, seven categories achieve 100% accuracy, while the average AUROC for object categories reaches 99.6%. This indicates that the proposed model exhibits strong robustness in anomaly detection across diverse object categories.