文献类型：期刊文献

英文题名：Dual-Teacher Network with SSIM Based Reverse Distillation for Anomaly Detection

作者：Li, Weihao[1]; Huang, Rongjin[1]; Wang, Zhanquan[1]

机构：[1] The College of Information Science and Engineering, East China University of Science and Technology, Shanghai, China

年份：2025

卷号：15043 LNCS

起止页码：266

外文期刊名：Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

收录：EI(收录号：20244817428925)

语种：英文

外文关键词：Anomaly detection - Benchmarking - Image segmentation - Personnel training - Teaching

摘要：There are many important applications for anomaly detection, such as industrial image anomaly detection, medical image anomaly detection etc. The teacher-student network (T-S) offers an effective method for identifying anomalies. Previous studies mainly utilize normal data or pseudo-anomalous for training, relying on cosine similarity for differentiation and usually involving only one guiding teacher. This approach limits the learning capacity of the student model, and solely concentrating on pixel-level differences might may hinder the model's grasp of overall anomaly structure and context, resulting in inadequate abnormal behavior recognition. To solve the problems, Dual-Teacher Network with SSIM based on Reverse Distillation (DSRD) is proposed, which combines the functionalities of two pre-trained teacher networks, namely T1 and T2, along with a denoising student encoder-decoder (S), a trainable one-class bottleneck embedding module (OCBE), and a segmentation network, all integrated within a unified framework. This comprehensive approach aims to leverage the strengths of each component to enhance outlier point detection and anomaly recognition capabilities. In the experiments on industrial inspection benchmark datasets, our method achieved state-of-the-art performance. The image-level AUC reached 99.41%, the pixel-level AUC reached 98.80%, the pixel-level average precision reached 79.34%, and the AUPRO reached 94.91%. ? The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025.