Class-Sensitive TPB-Guided Memory Refinement for Online Zero-Shot Anomaly Detection

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Zero-shot anomaly detection is crucial for industrial inspection, especially for new products lacking training data.
• CLIP-based methods offer generalization, and online memory improves adaptability but risks incorporating unreliable data.
• Existing methods struggle with weak or unstable visual categories due to memory update issues.

Purpose of the Study:

• To propose TSMR, a lightweight extension of RareCLIP for reliable online zero-shot anomaly detection.
• To enhance test-time memory evolution through a class-sensitive selective update strategy.
• To improve anomaly detection performance without altering the backbone or scoring pipeline.

Main Methods:

• TSMR employs a class-sensitive selective update strategy for test-time memory evolution.
• Key components include a confidence quantile gate, text-prior reliability check, and weak-class selective activation.
• A frame-level memory-update decision is made during online inference.

Main Results:

• TSMR significantly improves performance on the VisA dataset across image-level AUROC, pixel-level AUROC, and PRO metrics.
• It maintains competitive results on the MVTec AD dataset.
• Selective memory refinement proves especially beneficial for weak categories and is stable across different evaluation orders.

Conclusions:

• Reliable online memory evolution is an effective strategy for CLIP-based zero-shot anomaly detection.
• TSMR offers a robust approach to handling unreliable or ambiguous evidence in online updates.
• The proposed method enhances adaptability and performance in industrial inspection scenarios.