Unsupervised anomaly detection based on universal visual transformer and attention enhancement
WANG Zhen
ZHAI Ke
XUE Sai
BAI Shuang
Abstract：To address the common issues in existing unsupervised anomaly detection methods,such as insufficient feature extraction and the inability to effectively focus on anomalous regions,which lead to degraded detection performance,we propose an unsupervised anomaly detection method based on a general vision model and attention enhancement.First,the proposed method utilizes a pre-trained general vision model,the Vision Transformer(ViT),to extract features from input images.Second,to further enhance the model's focus on abnormal regions,we incorporate the Convolutional Block Attention Module(CBAM),which adaptively adjusts feature weights during the feature extraction stage to more precisely capture local anomalous information.Additionally,extensive experiments are conducted on the MVTec industrial dataset and a self-made cable anomaly dataset to comprehensively evaluate the detection performance of the proposed method.The experimental results demonstrate that the proposed method outperforms multiple state-of-the-art approaches in unsupervised anomaly detec-tion tasks.Specifically,on the cable anomaly dataset,the proposed method achieves an Image-wise AUROC(Image-wise Area Under ROC)and F1-Score of 88.1％and 80.8％,respectively,outper-forming the baseline Fastflow algorithm by 11.7％and 7.8％.
Keywords：anomaly detectionunsupervised detectionmachine visionvision transformerattention mechanism
Publication Date：2025-06-30
Online Publishing Date：2025-08-21(First online date of this platform, not the publication date of the document)
Pages：9( 14-22 )