Improved Deep Support Vector Data Description Model Using Feature Patching for Industrial Anomaly Detection

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Anomaly detection is critical for industrial quality control and operational efficiency.
• Deep Support Vector Data Description (SVDD) is a robust model for anomaly detection.
• Existing methods may not fully capture both macro-structural and local details in features.

Purpose of the Study:

• To propose an improved Deep SVDD model, Feature-Patching SVDD (FPSVDD), for unsupervised anomaly detection in industrial settings.
• To enhance sensitivity to anomalies by integrating a feature-patching technique.
• To improve real-time defect detection for manufacturing quality assurance.

Main Methods:

• Utilizing high-level features from ImageNet-trained networks.
• Implementing a feature-patching technique to split extracted features into smaller patches.
• Aggregating and concatenating feature patches for training with the Deep SVDD framework.
• Employing unsupervised learning for anomaly detection.

Main Results:

• The proposed FPSVDD model demonstrates superior performance compared to current mainstream approaches.
• Significant improvements in anomaly detection accuracy were observed on MvTec AD and CIFAR-10 datasets.
• The feature-patching method effectively captures both macro-structural and fine-grained local information.

Conclusions:

• FPSVDD offers enhanced anomaly detection capabilities for industrial applications.
• The model's ability to analyze feature patches improves defect detection sensitivity.
• FPSVDD is a valuable tool for real-time quality assurance in manufacturing.