DGIST Scholar: Unsupervised Anomaly Detection Using Self-Supervised Learning with Knowledge Distillation

Department of Robotics and Mechatronics Engineering Theses Master

Cited time in webofscience

Cited time in scopus

Unsupervised Anomaly Detection Using Self-Supervised Learning with Knowledge Distillation

Title: Unsupervised Anomaly Detection Using Self-Supervised Learning with Knowledge Distillation

Description: Anomaly Detection, Deep learning, Self-Supervised Learning, Knowledge Distillation

Table Of Contents: I. Introduction 1
II. Related Works 4
1 Reconstruction-based anomaly detection 4
2 Representation-based anomaly detection 5
III. Proposed Framework 7
1 Overview 7
2 Feature Extraction 8
3 CutPaste Augmentation 8
4 Knowledge Distillation with Contrastive Task 9
5 Anomaly Scoring Metric 12
6 Inference 12
IV. Experiments and Results 14
1 Experimental setup 14
1.1 Datasets 14
1.2 Implementation Details 15
1.3 Baselines 16
2 Comparison with State-of-the-Art Methods 17
2.1 MVTec-LOCO dataset 17
2.2 MVTec-AD dataset 18
2.3 CIFAR10 dataset 19
V. Ablation Study 21
1 Task Influence 21
2 Multi-Head Attention Influence 21
3 CutPaste Variants Influence 22
4 Contrastive Transformations influence 22
VI. Discussion 24
VII. Conclusion 25
References 26