# ReConPatch : Contrastive Patch Representation Learning for Industrial Anomaly Detection

**Authors:** Jeeho Hyun, Sangyun Kim, Giyoung Jeon, Seung Hwan Kim, Kyunghoon Bae, Byung Jun Kang  
**arXiv:** [2305.16713](https://arxiv.org/abs/2305.16713)  
**Date:** 2023-05-26  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2305.16713v1/

## Key points

- Accurately identifying product defects is crucial in industrial manufacturing
- Anomaly detection in this context poses challenges due to rarity of observations and unknown types of defects
- Existing methods utilize visual representations from natural image datasets and extract relevant features, but still face issues such as discrepancies between pre-trained features and target data or the need for carefully designed input augmentation specific to industrial datasets
- ReConPatch is a novel approach for anomaly detection in industrial settings
- ReConPatch constructs discriminative features by training a linear modulation attached to a pre-trained model
- It employs contrastive representation learning to collect and distribute features in a way that produces a target-oriented and easily separable representation
- ReConPatch utilizes two similarity measures (pairwise and contextual similarities) as pseudo-labels for contrastive learning, overcoming the challenge of absence of labeled pairs for training
- ReConPatch achieves robust anomaly detection performance without extensive input augmentation
- It achieves state-of-the-art results with an accuracy of 99.72% on the MVTec AD dataset

## Summary

In the field of industrial manufacturing, accurately identifying product defects such as incorrect parts, misaligned components, and damages is crucial. However, anomaly detection in this context poses challenges due to the rarity of observations and unknown types of defects. To address this difficulty, recent approaches have utilized visual representations from natural image datasets and extracted relevant features. However, these existing methods still face issues such as discrepancies between pre-trained features and target data or the need for carefully designed input augmentation specific to industrial datasets. To overcome these limitations, this paper introduces a novel approach called ReConPatch for anomaly detection in industrial settings. ReConPatch constructs discriminative features by training a linear modulation attached to a pre-trained model. It employs contrastive representation learning to collect and distribute features in a way that produces a target-oriented and easily separable representation. One challenge in contrastive learning is the absence of labeled pairs for training. To overcome this, ReConPatch utilizes two similarity measures: pairwise and contextual similarities between data representations serve as pseudo-labels for contrastive learning. Unlike previous methods, ReConPatch achieves robust anomaly detection performance without extensive input augmentation. In fact, it achieves state-of-the-art results with an accuracy of 99.72% on the widely used and challenging MVTec AD dataset. The authors of this paper are Jeeho Hyun, Sangyun Kim, Giyoung Jeon, Seung Hwan Kim, Kyunghoon Bae, and Byung Jun Kang.