Researchers from Capital Normal University and the School of Artificial Intelligence at Beijing University of Posts and Telecommunications have developed RealNet, a new feature reconstruction framework for industrial image anomaly detection. This approach addresses ongoing issues with generating diverse, realistic anomalies that align with natural distributions, as well as challenges around feature redundancy and pre-training bias.
Anomaly detection is a crucial aspect of industrial quality control and safety monitoring. Traditional methods are typically unsupervised and rely on normal data for training, falling into four categories: reconstruction-based, self-supervised learning, deep feature embedding, or one-class classification. These can struggle with effective reconstruction of anomalies, and suffer from feature redundancy and selection issues. Self-supervised methods, such as the Strength-controllable Diffusion Anomaly Synthesis (SDAS) incorporated in RealNet, allow for anomaly synthesis without labeled data, instead using normal images to simulate realistic anomalies.
The RealNet model incorporates SDAS, which generates anomalous images of varying strengths; Anomaly-aware Features Selection (AFS), which selects distinctive pre-trained features and reduces redundancy; and Reconstruction Residuals Selection (RRS), which adaptively selects distinctive anomalies. This combination allows the model to effectively use pre-trained Convolutional Neural Network features, drastically reducing redundancy and bias.
RealNet outperforms existing methods for image anomaly detection, demonstrating significant improvements in Image AU-ROC and Pixel AUROC in four benchmark datasets when compared to leading, reconstruction-based alternatives. Moreover, the synthetic anomalies RealNet generates closely align with the distribution of real anomalies, demonstrating the potential effectiveness of this approach in real-world scenarios.
RealNet introduces the Synthetic Industrial Anomaly Dataset (SIA), which further facilitates self-supervised anomaly detection. The results suggest that RealNet represents a cutting-edge contribution to the field and a versatile platform for future research, particularly on pre-trained feature reconstruction techniques. Moving forward, the researchers aim to extend their methods to tackle various real-world anomaly detection problems with increased proficiency and effectiveness.
The work is significant in the field of industrial quality control and safety monitoring, providing a sophisticated tool for anomaly detection. It highlights the importance of incorporating different methods and learning strategies, and underlines the potential of the pre-training phase in the anomaly detection process.
