Reconstruction from a distilled encoder with edge-based pseudo-anomaly for industrial anomaly detection

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Quality Control

Background:

• Supervised learning for industrial anomaly detection is hindered by limited anomaly samples.
• Existing reconstruction-based methods struggle with pseudo-anomaly quality and reconstruction accuracy.
• Reconstruction networks face challenges balancing image fidelity and anomaly localization.

Purpose of the Study:

• To propose RDEAD, an unsupervised anomaly detection framework addressing limitations in pseudo-anomaly generation and reconstruction.
• To enhance the accuracy and localization of anomaly detection in industrial settings.
• To improve the balance between image fidelity and precise anomaly reconstruction.

Main Methods:

• Developed an Edge-based pseudo-anomaly generation strategy (EPA) for realistic anomaly simulation.
• Introduced a distillation-based dual-encoder reconstruction network (YNet) for accurate anomaly reconstruction.
• Integrated an encoder feature fusion module (EFFM) within YNet to enhance detection performance.

Main Results:

• EPA generates pseudo-anomalies with improved shape similarity and localization.
• YNet effectively reconstructs anomalous regions while maintaining image fidelity.
• The RDEAD framework demonstrated significant effectiveness on multiple industrial datasets.

Conclusions:

• RDEAD offers a robust unsupervised approach for industrial anomaly detection.
• The proposed EPA and YNet components significantly advance reconstruction-based methods.
• The framework shows strong potential for real-world industrial quality control applications.