Fast Anomaly Detection for Vision-Based Industrial Inspection Using Cascades of Null Subspace PCA Detectors.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Anomaly detection is crucial for quality control in automated manufacturing.
• Existing methods often require computationally intensive models and high-end hardware.
• There is a need for efficient anomaly detection solutions for resource-constrained settings.

Purpose of the Study:

• To develop a novel, computationally efficient anomaly detection framework for industrial imaging.
• To leverage lightweight convolutional neural network (CNN) features and a unique PCA approach for enhanced sensitivity.
• To provide a practical solution for anomaly detection without high-end hardware requirements.

Main Methods:

• Utilized MobileNetV2 as a lightweight CNN backbone for feature extraction.
• Developed a PCA-based anomaly detection module focusing on near-zero variance features, exploiting the approximate null space.
• Implemented a cascaded multi-stage decision process using local features from each CNN layer independently.

Main Results:

• Achieved superior anomaly detection performance on MVTec (99.4% AUROC) and VisA (91.7% AUROC) benchmark datasets.
• Demonstrated significant computational efficiency compared to existing methods.
• Validated the effectiveness of exploiting low-variance features and cascaded decision-making.

Conclusions:

• The proposed framework offers a compelling solution for practical anomaly detection in industrial settings.
• The method achieves competitive accuracy with minimal hardware resources.
• This approach enhances sensitivity to anomalies while reducing computational complexity.