# Towards Total Recall in Industrial Anomaly Detection

**Authors:** Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, Peter Gehler  
**arXiv:** [2106.08265](https://arxiv.org/abs/2106.08265)  
**Date:** 2021-06-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2106.08265v1/

## Key points

- The ability to identify defective parts in large-scale industrial manufacturing is crucial.
- The cold-start problem is a significant challenge, which involves training a model using only non-defective example images.
- Researchers have combined embeddings from ImageNet models with outlier detection models to address this challenge, resulting in the best performing approaches thus far.
- PatchCore is a novel method that utilizes a memory bank of highly representative nominal patch-features.
- PatchCore achieves state-of-the-art performance for both anomaly detection and localization, with competitive inference times.
- On the MVTec AD dataset, PatchCore achieves an impressive image-level anomaly detection AUROC score of 99.1%, surpassing its closest competitor by more than half in terms of error reduction.
- Competitive results are reported on two other datasets, highlighting the versatility and generalizability of PatchCore's approach.
- PatchCore exhibits promising results even when dealing with limited samples, indicating its potential applicability in scenarios where data availability may be restricted.
- This research significantly advances the field of industrial anomaly detection by proposing an innovative solution that outperforms existing methods while maintaining efficient computational performance.

## Summary

In the field of large-scale industrial manufacturing, the ability to identify defective parts is crucial. A significant challenge in this area is the cold-start problem, which involves training a model using only non-defective example images. Researchers have combined embeddings from ImageNet models with outlier detection models to address this challenge, resulting in the best performing approaches thus far. Building upon this line of work, this paper introduces PatchCore, a novel method that utilizes a memory bank of highly representative nominal patch-features. PatchCore not only achieves state-of-the-art performance for both anomaly detection and localization but also offers competitive inference times. The effectiveness of PatchCore is demonstrated on the widely used MVTec AD dataset, where it achieves an impressive image-level anomaly detection AUROC score of 99.1%, surpassing its closest competitor by more than half in terms of error reduction. Additionally, competitive results are reported on two other datasets, highlighting the versatility and generalizability of PatchCore's approach. Furthermore, PatchCore exhibits promising results even when dealing with limited samples indicating its potential applicability in scenarios where data availability may be restricted. Overall, this research significantly advances the field of industrial anomaly detection by proposing an innovative solution that outperforms existing methods while maintaining efficient computational performance.