LA-EAD: Simple and Effective Methods for Improving Logical Anomaly Detection Capability.

Area of Science:

• Intelligent Manufacturing
• Computer Vision
• Machine Learning

Background:

• Automated product quality inspection relies heavily on image anomaly detection.
• Existing methods excel at detecting local structural anomalies but struggle with global logical anomalies.
• Logical anomalies require models capable of extracting global context features.

Purpose of the Study:

• To develop a lightweight anomaly detection framework for intelligent manufacturing.
• To improve the detection of both structural and logical anomalies.
• To balance the detection capabilities for diverse anomaly types.

Main Methods:

• Proposed a framework integrating reconstruction difference constraint (RDC) and a logical anomaly detection module, building upon EfficientAD.
• RDC enhances fine-grained reconstruction consistency, mitigating false detections.
• A logical anomaly detection module extracts and aggregates global context features for anomaly scoring.

Main Results:

• Achieved 94.2 AU-ROC for logical anomaly detection on MVTec LOCO.
• Maintained strong structural anomaly detection performance with 98.4 AU-ROC on MVTec AD.
• Demonstrated a state-of-the-art balance between detecting structural and logical anomalies compared to baselines.

Conclusions:

• The proposed framework effectively addresses the challenge of detecting both structural and logical anomalies.
• The integration of RDC and a dedicated logical anomaly module significantly improves detection accuracy.
• This method offers a balanced and high-performing solution for automated quality inspection in intelligent manufacturing.