Open sy00n opened 8 months ago
E: Pre-trained teacher encoder (extracts multi-scale representation)
Trainable one-class bottle-neck embedding module
Student decoder D (restore the features from the bottleneck embedding)
T-S 간 abnormality score가 낮다는 것은 높은 abnormality score를 의미함.
Heterogeneous encoder, decoder 구조와 knowledge distillation 순서가 anomaly에 대한 discrepant representation에 크게 기여함을 주장함.
Trainable OCBE 모듈이 mult-scale pattern을 downstream normal representation reconstruction을 위한 low-dimensional space로 통합함.
Abstract
Introduction
E: Pre-trained teacher encoder (extracts multi-scale representation)
Trainable one-class bottle-neck embedding module
Student decoder D (restore the features from the bottleneck embedding)
T-S 간 abnormality score가 낮다는 것은 높은 abnormality score를 의미함.
Heterogeneous encoder, decoder 구조와 knowledge distillation 순서가 anomaly에 대한 discrepant representation에 크게 기여함을 주장함.
Trainable OCBE 모듈이 mult-scale pattern을 downstream normal representation reconstruction을 위한 low-dimensional space로 통합함.
Reverse Distillation