Implementing Multi-Level Features in a Student-Teacher Network for Anomaly Detection

Abstract

Anomaly detection is an open and challenging problem that aims to detect anomaly in future samples. In this study, we explore a simple but effective solution that utilizes multi-level feature combination in a student-teacher network to improve the prediction result. Our approach combines low-level, middle-level, and high-level features extracted from ResNet18 to capture a range of features from different layers of the network. Through the use of a student-teacher network, we select the best possible generated features from ResNet18 to enhance the prediction performance. Our results demonstrate that combining features from different levels of the network enhances the model's ability to learn and recognize anomalous patterns, and thus improves the accuracy of anomaly detection. Our proposed student-teacher network with ResNet18 backbone achieves a prediction score of 92.80% and 96.90% for Image AUC and Pixel AUC respectively.