EUAD : end-to-end unsupervised anomaly detection based on few normal data

Abstract

In recent years, unsupervised anomaly detection has attracted significant attention in the field, as it relies solely on unlabeled samples for training while achieving good performance. Most unsupervised models either adopt anomaly simulation to generate labeled samples or use memory banks to store normal features for comparison. However, existing simulation strategies mainly focus on object-type anomalies, with insufficient attention to fine-grained texture-type ones. Meanwhile, memory bank approaches whether by storing all data or using random sampling lead to excessive memory overhead and lack representativeness in the sampled data. To address these issues, we propose an end-to-end unsupervised anomaly detection framework (EUAD). Specifically, EUAD simulates a wider variety of realistic texture defects by expanding the texture anomaly library to improve the accuracy of texture anomaly detection. Then, EUAD obtains a representative feature set through feature clustering to improve the effectiveness and reliability of sample storage. Finally, we propose a new feature fusion module that integrates multi-scale input features and memory bank features, which includes the Convolutional Block Attention Module (CBAM) and Pixel-based Multi-scale Fusion Module (PMFM). Leveraging this module, EUAD fully utilizes limited normal features to suppress irrelevant information and highlight anomalies, ultimately improving detection sensitivity to anomalies of different sizes. Extensive experimental results on public datasets show that the proposed EUAD in this paper outperforms other anomaly detection models.