Student-Teacher Anomaly Detection Considering Knowledge Consistency between Layer Groups

Student-teacher networks have been widely used for anomaly detection, which is often addressed as a one-class classification task. The mainstream idea is to calculate the loss of multiple feature maps between the student network and the teacher network independently without considering their relevance to detect anomalies. In this paper, we introduce a knowledge consistency loss into the student-teacher framework for further improving the performance based on the observation that anomaly scores obtained between adjacent layer groups should be spatially consistent. Evaluational experiments on a publicly available benchmark confirmed that our proposal can improve pixel-level anomaly detection when the anomaly score map is calculated from the feature map in the highest resolution.