Visual defect obfuscation based self-supervised anomaly detection.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Quality Control

Background:

• Unsupervised anomaly detection (UAD) is crucial for manufacturing quality control due to limited anomaly data.
• Current UAD methods struggle with unseen anomalies and reconstruction accuracy.
• Reconstruction-by-inpainting methods show promise but face challenges with masking strategies.

Purpose of the Study:

• To develop a novel reconstruction-by-inpainting method for enhanced unsupervised anomaly detection.
• To address limitations of existing methods, including inference time, output consistency, and reconstruction accuracy.
• To propose a practical and deployable solution for industrial anomaly detection.

Main Methods:

• Introduced Excision And Recovery (EAR), a novel reconstruction-by-inpainting approach.
• Utilized single deterministic masking based on ImageNet pre-trained DINO-ViT.
• Incorporated visual obfuscation (mosaicing) for hint-providing during reconstruction.

Main Results:

• Deterministic masking effectively identified and isolated suspected defective regions.
• Resolved issues of time-consuming inference and random masking inconsistency.
• Mosaicing improved normal pattern reconstruction accuracy compared to binary masking.
• Achieved high performance without altering the underlying model structure.

Conclusions:

• EAR significantly enhances unsupervised anomaly detection performance in industrial settings.
• The method offers a practical solution by improving speed, consistency, and accuracy.
• Future work will focus on evaluating EAR's applicability in diverse manufacturing environments.