Abnormal Discrepancy-Guided Knowledge Distillation for Image Anomaly Detection

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Knowledge distillation is crucial for image anomaly detection, but current methods struggle with feature alignment loss causing reconstruction errors.
• Existing techniques often confuse reconstruction errors with actual anomalies, limiting detection accuracy.

Purpose of the Study:

• To propose an enhanced knowledge distillation method, abnormal discrepancy-guided knowledge distillation (DiffKD), for superior image anomaly detection.
• To address the limitations of feature alignment loss in current methods by introducing differential guidance for student feature reconstruction.

Main Methods:

• DiffKD employs channel-level discrepancy masks to guide student feature reconstruction, using normal features as supervision and abnormal features as constraints.
• The method integrates a knowledge distillation network for feature reconstruction and a segmentation network for anomaly localization.
• Utilizes prior and synthetic anomaly samples for real-time training data of anomalous samples.

Main Results:

• DiffKD achieves 80.7% average precision (AP) and 81.9% instance-level average precision (IAP) on the MVTec AD benchmark.
• Demonstrates competitive performance against representative methods under the same evaluation protocol.
• Validated effectiveness and generalizability on SUT-Crack and MVTec AD benchmarks.

Conclusions:

• The proposed DiffKD method significantly improves image anomaly detection accuracy.
• DiffKD shows promise for enabling real-time, automated anomaly detection in practical applications.