TLDR: A new research paper introduces Normal-Abnormal Guided Generalist Anomaly Detection (NAGL), a novel framework that significantly improves anomaly detection across diverse domains. Unlike previous methods that rely solely on normal samples, NAGL leverages both normal and a small number of anomalous samples as references. It employs Residual Mining to extract abnormal patterns and Anomaly Feature Learning to identify instance-aware anomalies, leading to more accurate, efficient, and generalizable detection performance across industrial and medical datasets.
Anomaly detection, a critical task in fields ranging from industrial quality control to medical diagnosis, aims to identify unusual patterns that deviate from expected behavior. Traditionally, models for anomaly detection are trained and tested within a single domain, often relying solely on examples of what is considered ‘normal’. However, real-world scenarios frequently present challenges like data scarcity and privacy concerns, making it difficult to train a model directly on a new target domain.
This is where Generalist Anomaly Detection (GAD) comes into play. GAD seeks to train a single, unified model on an original domain that can then effectively detect anomalies in entirely new, unseen target domains. While existing GAD methods have made progress, they primarily use only normal samples as a reference point. This overlooks a crucial piece of information: anomalous samples, which are often available in small quantities in real-world settings (e.g., a few defective parts or diagnosed disease cases).
A new research paper, titled Normal-Abnormal Guided Generalist Anomaly Detection, introduces a more practical and effective approach. Authored by Yuexin Wang, Xiaolei Wang, Yizheng Gong, and Jimin Xiao, this work proposes leveraging both normal and anomalous samples as references to guide anomaly detection across diverse domains. This innovative strategy is the first of its kind in generalist anomaly detection.
The core of their approach is the Normal-Abnormal Generalist Learning (NAGL) framework, which consists of two main components: Residual Mining (RM) and Anomaly Feature Learning (AFL). Think of it this way: RM is designed to extract the unique patterns of anomalies from the differences between normal and abnormal reference examples. This helps the model build a transferable understanding of what an anomaly ‘looks like’. Then, AFL takes this understanding and adaptively learns to identify specific anomalous features in new images by comparing them to these learned abnormal patterns.
This dual-guidance system allows the model to capture the similarities and differences between a query image and both normal and abnormal reference samples. The researchers found that directly using abnormal samples in traditional methods could lead to ‘false activations,’ where normal patterns are mistakenly identified as anomalies. The NAGL framework effectively mitigates this issue by carefully mining abnormal variations in a ‘residual space’ to guide detection, leading to more accurate results.
Extensive experiments conducted across multiple benchmarks, including industrial datasets like MVTecAD and VisA, and even medical datasets like BraTS, demonstrate the significant advantages of this new method. The NAGL framework consistently outperforms existing GAD approaches, showing superior generalization capabilities across different industrial scenarios and even when transferring from industrial to medical applications. Remarkably, incorporating just one abnormal reference sample yielded substantial performance gains.
Beyond accuracy, the NAGL framework also proves to be highly efficient. It boasts a significantly smaller model size, requires less training time, and achieves faster inference speeds compared to previous generalist anomaly detection methods. For instance, it is reported to be 14 times faster than one prominent existing method and 2 times faster than another in terms of inference speed, making it highly suitable for practical, real-world deployment.
Also Read:
- Making AI Explanations More Reliable with Activation-Deactivation
- Bridging Disparate Data with Indirect Attention
This research marks a significant step forward in anomaly detection, particularly for scenarios where a small number of anomalous examples can be provided. By effectively utilizing this often-overlooked information, the NAGL framework opens new avenues for more robust and efficient anomaly detection systems in various critical applications.
