# CFA: Coupled-hypersphere-based Feature Adaptation for Target-Oriented Anomaly Localization

**Authors:** Sungwook Lee, Seunghyun Lee, Byung Cheol Song  
**arXiv:** [2206.04325](https://arxiv.org/abs/2206.04325)  
**Date:** 2022-06-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2206.04325v1/

## Key points

- The paper addresses the problem of distinguishing normal and abnormal features within a dataset
- Previous studies have overestimated the normality of abnormal features by not adapting to the target dataset
- The authors propose a method called Coupled-hypersphere-based Feature Adaptation (CFA)
- CFA consists of a learnable patch descriptor and a scalable memory bank
- The learnable patch descriptor enables precise discrimination between normal and abnormal features
- The scalable memory bank improves anomaly detection performance
- CFA outperforms existing approaches with an AUROC score of 99.5% in anomaly detection and 98.5% in anomaly localization on the MVTec AD benchmark dataset
- Adapting pre-trained CNNs to the target dataset is important for accurate anomaly localization
- Code for implementing CFA is available on GitHub at https://github.com/sungwool/CFA_for_anomaly_localization

## Summary

The paper titled "CFA: Coupled-hypersphere-based Feature Adaptation for Target-Oriented Anomaly Localization" by Sungwook Lee, Seunghyun Lee, and Byung Cheol Song addresses the problem of accurately distinguishing normal and abnormal features within a dataset. Previous studies have focused on approximating the distribution of normal features without adapting to the target dataset, leading to an overestimation of the normality of abnormal features. To overcome this limitation, the authors propose a method called Coupled-hypersphere-based Feature Adaptation (CFA). CFA consists of two main components: a learnable patch descriptor and a scalable memory bank. The learnable patch descriptor is responsible for learning and embedding target-oriented features that enable precise discrimination between normal and abnormal features. The scalable memory bank is independent of the size of the target dataset and plays a crucial role in improving anomaly detection performance. The authors evaluate their proposed method on the MVTec AD benchmark dataset and compare it with previous methods quantitatively and qualitatively. The results demonstrate that CFA outperforms existing approaches with an AUROC score of 99.5% in anomaly detection and 98.5% in anomaly localization. Additionally, this paper highlights the importance of adapting pre-trained CNNs to the target dataset for accurate anomaly localization. The code for implementing CFA is publicly available on GitHub at https://github.com/sungwool/CFA_for_anomaly_localization. Overall, CFA presents an effective solution for anomaly localization by leveraging adapted features through transfer learning and utilizing a scalable memory bank.