Anomaly detection plays a critical role in various industries for quality control and safety monitoring. The common methods of anomaly detection involve using self-supervised feature reconstruction. However, these techniques are often challenged by the need to create diverse and realistic anomaly samples while reducing feature redundancy and eliminating pre-training bias.
Researchers from the College of Information and Engineering at Capital Normal University and the School of Artificial Intelligence at Beijing University of Posts and Telecommunications have developed RealNet to address these issues. RealNet is a feature reconstruction framework that brings together existing methods to enhance anomaly detection while significantly lowering redundancy and bias. The framework comprises three key components: Strength-controllable Diffusion Anomaly Synthesis (SDAS), Anomaly-aware Features Selection (AFS), and Reconstruction Residuals Selection (RRS).
The SDAS component of RealNet creates various realistic anomalies that align with natural distributions. The AFS component helps in reducing cost and redundancy by selecting only the relevant pre-trained features, while RRS adaptively chooses the most effective residuals for anomaly identification.
The RealNet framework has been shown to outperform existing methods on benchmark datasets, while also introducing the Synthetic Industrial Anomaly Dataset (SIA) for improving anomaly synthesis in self-supervised detection methods. The researchers focused on self-supervised learning and reconstruction methods, as these techniques are central to the RealNet framework. SDAS, in particular, provides a means of synthesizing realistic anomalies without the need for labelled data.
An evaluation of RealNet’s performance using AU-ROC metrics showed a significant improvement in comparison to current state-of-the-art methods. The results confirmed that RealNet outperforms alternative methods, including PatchCore, SimpleNet, and FastFlow. An assessment of anomaly image quality generated by RealNet’s Frechet Inception Distance (FID) confirmed the synthetic anomaly images closely mimic real anomaly images.
The RealNet framework is a leap forward in self-supervised anomaly detection. Its three core components make it possible to leverage large-scale pre-trained models efficiently while maintaining computational efficiency. This innovative technology offers a promising platform for future research in anomaly detection, particularly in pre-trained feature reconstruction techniques. RealNet has demonstrated effective capabilities to tackle a variety of real-world anomaly detection scenarios.
You can find out more about RealNet by checking out their research paper and Github. To stay updated on their latest developments, follow them on Twitter, join their Discord Channel, Telegram Channel, or LinkedIn Group. If you are interested in their work, you might consider signing up for their newsletter.