A HIERARCHICAL-STRUCTURED ANOMALY DETECTION METHOD

Anomaly detection and localization play a key role in industrial pipelines. Most of the current methods only use a single type of features for anomaly detection and cannot achieve stable performance. To this end, a new hierarchical feature fusion framework is developed, that is, a structural feature fusion module is designed by jointly reconstructing high-level features and low-level features. Meanwhile, to address the problem of excessive generalization, the framework introduced an attention mechanism based on the Cutpaste method. For the two tasks of anomaly localization and anomaly detection, the AUROC indicators of this framework on the MVTec dataset reached 0.979 and 0.987, respectively, which was a significant improvement compared with the existing mainstream anomaly detection algorithms.