FR-PatchCore: An Industrial Anomaly Detection Method for Improving Generalization.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Quality Control

Background:

• Self-supervised anomaly detection methods are advancing rapidly.
• PatchCore is a state-of-the-art algorithm on the MVTec AD benchmark, known for efficiency and reduced need for labeled data.
• PatchCore struggles with anomalies in samples exhibiting positional variations like rotation or misalignment.

Purpose of the Study:

• To introduce Feature-Level Registration PatchCore (FR-PatchCore), an extension of PatchCore.
• To overcome the limitations of standard PatchCore in detecting anomalies with positional variations.
• To improve the accuracy and generalizability of industrial anomaly detection.

Main Methods:

• FR-PatchCore constructs and continuously updates a feature matrix in its memory bank.
• Optimal negative cosine similarity loss is employed for feature matrix updates.
• A novel method for computing the mask threshold is proposed for accurate anomaly segmentation.

Main Results:

• FR-PatchCore achieved a high image-level anomaly detection AUROC score of 98.81% on the MVTec AD benchmark.
• The proposed method effectively handles positional variations in samples.
• The novel mask thresholding method allows for precise partitioning of anomalous regions.

Conclusions:

• FR-PatchCore significantly enhances anomaly detection accuracy, especially in scenarios with positional variations.
• The method demonstrates high generalizability and strong potential for real-world industrial anomaly detection.
• FR-PatchCore offers a robust solution for quality control in manufacturing settings.