# What makes a good data augmentation for few-shot unsupervised image anomaly detection?

**Authors:** Shuheng Zhang, Lingrui Zhang, Guoyang Xie, Jiaqi Liu, Hua Yan, Jinbao Wang, Feng Zheng, Yaochu Jin  
**arXiv:** [2304.03294](https://arxiv.org/abs/2304.03294)  
**Date:** 2023-04-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.03294v1/

## Key points

- The paper explores the effectiveness of data augmentation methods for unsupervised anomaly detection in industrial applications
- Positive samples are often limited due to commercial competition and sample collection difficulties
- Six data augmentation methods are used: Flip, Scale, Translate, Color Jitter, Perspective, and Mixed Data Augmentation
- Experimental results show that there is no universal winner among data augmentation methods for all IAD algorithms
- Combining multiple data augmentation methods does not necessarily yield further improvements in the accuracy of anomaly detection; however, it can achieve excellent results on specific methods
- The study identifies the best single data augmentations for each category of IAD algorithm in each dataset
- Provides useful guidance on selecting appropriate data augmentation methods for different requirements in IAD
- Highlights the importance of carefully selecting an appropriate combination of techniques based on the specific application context rather than relying solely on a single approach.

## Summary

This paper explores the effectiveness of data augmentation methods for unsupervised anomaly detection in industrial applications, where positive samples are often limited due to commercial competition and sample collection difficulties. The study investigates how to select and apply data augmentation methods effectively for unsupervised anomaly detection. The impact of various data augmentation techniques on different anomaly detection algorithms is systematically examined through experiments using three datasets: MVtec AD, MTD, and BTAD. Six data augmentation methods are used: Flip, Scale, Translate, Color Jitter, Perspective, and Mixed Data Augmentation. For each augmentation method, one training image is augmented into four images. The experimental results show that there is no universal winner among data augmentation methods for all IAD (industrial image anomaly detection) algorithms. The performance of different IAD algorithms is not significantly affected by the specific data augmentation method employed. Combining multiple data augmentation methods does not necessarily yield further improvements in the accuracy of anomaly detection; however, it can achieve excellent results on specific methods. As for Memory Bank-based methods, mixed data augmentation significantly outperforms single data augmentations for image-level metrics. The study also identifies the best single data augmentations for each category of IAD algorithm in each dataset. For example, Flip is effective for Texture-based algorithms on MVtec AD dataset while Color Jitter works well with Deep Learning-based algorithms on MTD dataset. In conclusion, this research provides useful guidance on selecting appropriate data augmentation methods for different requirements in IAD and highlights the importance of carefully selecting an appropriate combination of techniques based on the specific application context rather than relying solely on a single approach.