Multi-Scale U-Shape Transformer Network for Unsupervised Image Anomaly Detection

Currently, the field of image anomaly detection faces significant challenges, primarily due to data scarcity and insufficient model performance. The lack of abnormal samples has made unsupervised methods a research priority; however, the detection performance of unsupervised algorithms typically falls short of that of supervised algorithms. To address this issue, this paper proposes a multi-scale transformer network based on pretrained network features and a reconstruction framework. This approach extracts deep features across multiple levels and performs reconstruction at the feature layer to accurately localize anomalies. Compared to traditional reconstruction-based anomaly detection methods, our approach leverages pretrained feature representations, thus avoiding the generalization issues caused by relying solely on normal samples. Additionally, we introduce a cross-attention module embedded within skip connections, which filters out non-semantic features and enables more refined spatial restoration and feature representation, further reducing the likelihood of overgeneralization and enhancing prediction accuracy. The performance of this method on the industrial dataset MVtecAD and medical datasets demonstrates its strong generalization ability and effective anomaly detection performance.