Adapting Generic Features to A Specific Task: A Large Discrepancy Knowledge Distillation for Image Anomaly Detection
Chenkai Zhang (Zhejiang University),* Tianqi Du (Zhejiang University), Yueming Wang (Zhejiang University)
The 34th British Machine Vision Conference
Abstract
Anomaly detection is a challenging task due to the lack of data on unexpected anomalies. Recent approaches using Knowledge Distillation (KD) between Teacher-Student (T-S) models have shown great potential for anomaly detection. These techniques use pre-trained models on natural images as the teacher model. However, for industrial images, defects typically occur in a small region, while the global semantics of the anomaly image remain similar to normal images. This situation results in generic features being unable to capture defects well, leading to a loss of discriminability in detecting anomalies. This paper proposes a way to improve this situation by applying learnable feature mappings to adapt the generic features for the data-specific task. Additionally, a novel angular margin loss is introduced to improve the regular training loss of knowledge distillation and ensure larger discrepancies between T-S models on anomalies. Extensive experiments show that the proposed feature mappings and angular loss can effectively improve the feature discriminability for anomaly detection and help state-of-the-art KD-based methods achieve better detection performance.Video
Citation
@inproceedings{Zhang_2023_BMVC,
author = {Chenkai Zhang and Tianqi Du and Yueming Wang},
title = {Adapting Generic Features to A Specific Task: A Large Discrepancy Knowledge Distillation for Image Anomaly Detection},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://papers.bmvc2023.org/0376.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by