PatchCore: Paper Overview and Japanese Translation PDF

Hello, this is Seong-Hun Choe from TOMOMI RESEARCH.
This time, I would like to introduce a notable research paper published in 2022, "Towards Total Recall in Industrial Anomaly Detection." The study proposes a new method called "PatchCore" that significantly advances anomaly detection in industrial image data.

Source paper
https://arxiv.org/abs/2106.08265

What Is PatchCore?

PatchCore is an anomaly detection method that addresses the "cold-start" problem, detecting anomalies mainly by using "normal images only." The method has the following characteristics:

• High accuracy: it records a remarkable AUROC score of up to 99.6% on the widely used MVTec AD benchmark.
• Efficiency: fast inference, with lower memory usage than conventional methods.
• Flexibility: it can detect a wide range of anomalies, from small scratches to large structural changes.

How It Differs from Conventional Methods

Conventional anomaly detection methods used features pre-trained on ImageNet, but many of them were not optimized for industrial data and had issues with accuracy and efficiency. PatchCore improves on this in the following ways:

• Local patch features: uses features from intermediate layers to retain spatial context.
  

• Coreset subsampling: reduces redundant data, greatly saving memory and computation.
  

• Anomaly scoring: detects anomalies patch by patch, enabling fine-grained anomaly localization.

Practicality and Outlook

The performance of PatchCore has been verified on the MVTec AD benchmark and on other datasets (e.g., MTD and mSTC), showing practical inference speed and high accuracy even with small amounts of data. On the other hand, because it relies on ImageNet pre-trained features, further adaptation may be needed in specific industrial environments. Going forward, combining it with pre-training specialized for industrial settings is expected to improve performance further.

Japanese Translation of the Paper Available as a PDF

We have prepared a Japanese translation of the paper introduced here in PDF format, for readers who prefer Japanese. You can download it from the link below.

Download the PDF (in Japanese)
patchcore_Japanese

PatchCore is a technology with the potential to transform the field of industrial anomaly detection, and it will likely see growing adoption on real production floors and in inspection systems. Drawing on research results like this, we want to accelerate our own technology development even further!