A generalized defect-data-free defect inspection method based on image reconstruction and anomaly detection.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Traditional supervised methods for defect detection require extensive defect-specific training data.
• This limitation hinders generalization across diverse product types and real-world industrial scenarios.

Purpose of the Study:

• To develop a novel, unsupervised framework for industrial defect inspection.
• To enable accurate and efficient anomaly detection without prior knowledge of defect types.

Main Methods:

• The framework utilizes hierarchical image reconstruction modules.
• A self-attention mechanism is incorporated for enhanced feature learning.
• Anomaly detection is performed based on reconstruction errors.

Main Results:

• Achieved an average precision of 97.83% on the MVTec AD 2D dataset.
• Outperformed U-Net by 11.1% and U-Transformer by 12.9% in accuracy.
• Reached a model inference speed of 24.1 FPS, 48.1% faster than U-Transformer models.

Conclusions:

• The proposed framework offers a robust solution for real-time industrial defect inspection.
• Demonstrates superior performance in both detection accuracy and inference speed.
• Highlights the potential of unsupervised learning for versatile anomaly detection.