PA-CLIP: Enhancing Zero-Shot Anomaly Detection through Pseudo-Anomaly Awareness

TL;DR

PA-CLIP tackles robust zero-shot industrial anomaly detection by introducing pseudo-anomaly awareness through a tripartite CLIP-based framework. It combines Multi-Scale Feature Aggregation, Pseudo-Anomaly-Aware Memory, and Pseudo-Anomaly-Aware Decision to disentangle background variations from true defects using dual memory banks and adaptive fusion. The method achieves state-of-the-art performance on MVTec AD and VisA, with strong classification and segmentation gains and demonstrated generalization to other CLIP-based approaches. This work highlights the practical potential of zero-shot defect detection with unlabeled test data and category-name prompts in real-world industrial settings.

Abstract

In industrial anomaly detection (IAD), accurately identifying defects amidst diverse anomalies and under varying imaging conditions remains a significant challenge. Traditional approaches often struggle with high false-positive rates, frequently misclassifying normal shadows and surface deformations as defects, an issue that becomes particularly pronounced in products with complex and intricate surface features. To address these challenges, we introduce PA-CLIP, a zero-shot anomaly detection method that reduces background noise and enhances defect detection through a pseudo-anomaly-based framework. The proposed method integrates a multiscale feature aggregation strategy for capturing detailed global and local information, two memory banks for distinguishing background information, including normal patterns and pseudo-anomalies, from true anomaly features, and a decision-making module designed to minimize false positives caused by environmental variations while maintaining high defect sensitivity. Demonstrated on the MVTec AD and VisA datasets, PA-CLIP outperforms existing zero-shot methods, providing a robust solution for industrial defect detection.

Figures (3)

• Figure 1: Left: Example images from two categories are shown: anomalous images and entirely normal images. Green indicates normal regions, orange indicates pseudo-anomalous regions, and red indicates anomalous regions. The heatmap is generated by a SOTA model. Top Right: This subfigure illustrates the feature distribution after dimensionality reduction. Green points represent normal features, orange points represent pseudo-anomalous features, and red points represent true anomalous features. Bottom Left: Our model outperforms others in distinguishing pseudo-anomalies.
• Figure 2: Overview of PA-CLIP. It involves a three-stage process: (a) Multi-Scale Aggregation, (b) Pseudo-Anomaly-Aware Memory, (c) Pseudo-Anomaly-Aware Decision.
• Figure 3: Performance comparison of anomaly detection methods, including AUROC metrics and segmentation visualization.