[CVPR2022] Anomaly Detection via Reverse Distillation from One-Class Embedding
CVPR 2022
Link: [2201.10703] Anomaly Detection via Reverse Distillation from One-Class Embedding (arxiv.org)
Code: https://github.com/Merenguelkl/Reverse_Disstilation
1. Intro
![https://i-blog.csdnimg.cn/blog_migrate/807da91ba10de4c6c0b5b9c58f439fa3.jpeg](https://i-blog.csdnimg.cn/blog_migrate/807da91ba10de4c6c0b5b9c58f439fa3.jpeg)
提出一个新颖的知识蒸馏框架:Teacher Encoder + Student Decoder,并使用逆向蒸馏(Reverse Distillation),引入一个可训练的 one-class bottleneck embedding (OCBE) 模块作为bottleneck,该模块可以保留用于还原正常样本特征的信息。OCBE包含了multi-scale feature fusion block (MFF)