Reproduce “ProtoAnomalyNCD: Prototype Learning for Multi-class Novel Anomaly Discovery in Industrial Scenarios” on OpenResearch

About this paper

Botong Zhao, Qijun Shi, Shujing Lyu, Yue Lu

ProtoAnomalyNCD is a prototype-learning framework developed by researchers at East China Normal University that automatically discovers and classifies multiple types of visual anomalies in industrial inspection, including previously unseen categories. The framework leverages anomaly maps and object localization to guide attention, dynamically estimates the number of novel anomaly classes, and extends to detect out-of-distribution samples, outperforming existing anomaly clustering and NCD methods on industrial datasets.

Key insights

• Explicitly guiding the Vision Transformer's attention with anomaly maps and foreground object masks significantly improves the model's ability to focus on and extract features from subtle, fine-grained industrial anomalies.
• Prototype-based probabilistic modeling within a shared hyperspherical feature space, combined with a dual-view self-supervised learning setup, effectively facilitates the discovery and separation of multiple novel anomaly classes without requiring their labels.
• Dynamically estimating the number of novel classes by analyzing the consistency and performance of base class classification provides a robust solution to a major limitation of traditional Novel Class Discovery methods in open-set industrial scenarios.

Summary by alphaXiv.