In industrial image anomaly detection, self-supervised feature reconstruction methods show promise but still grapple with challenges such as generating realistic and diverse anomaly samples while mitigating feature redundancy and pre-training bias. Synthetic anomalies lack diversity and realism, hindering model generalization. Meanwhile, feature reconstruction-based detection, though simple, needs to improve with high computational demands and requires more effective feature selection. Recent studies emphasize the importance of feature selection, urging a unified approach to advance anomaly detection, which is crucial in industrial quality control and safety monitoring.
Researchers from the College of Information and Engineering, Capital Normal University, and School of Artificial Intelligence, Beijing University of Posts and Telecommunications have developed RealNet, a feature reconstruction framework incorporating Strength-controllable Diffusion Anomaly Synthesis (SDAS) that generates diverse, realistic anomalies aligned with natural distributions, Anomaly-aware Features Selection (AFS), and Reconstruction Residuals Selection (RRS). RealNet enhances anomaly detection by efficiently utilizing pre-trained CNN features, reducing redundancy and bias. It introduces SDAS for realistic anomaly synthesis, AFS for feature selection, and RRS for adaptive residual selection. RealNet outperforms existing methods on benchmark datasets and introduces the Synthetic Industrial Anomaly Dataset (SIA) for anomaly synthesis, facilitating self-supervised detection methods.
Unsupervised anomaly detection methods rely solely on normal data for training, falling into four categories: reconstruction-based, self-supervised learning, deep feature embedding, and one-class classification. The study focuses on reconstruction and self-supervised learning methods, which are crucial for the RealNet framework. While reconstruction methods struggle with effectively reconstructing anomalies, recent studies emphasize anomaly detection through pre-trained feature reconstruction. However, challenges persist in feature redundancy and selection across different anomaly categories. In contrast, self-supervised methods like SDAS enable realistic anomaly synthesis without labeled data, offering control over anomaly strengths solely using normal images.
RealNet is a framework for anomaly detection consisting of SDAS, AFS, and RRS. SDAS generates anomalous images with varying strengths, mimicking real anomalies. AFS selects discriminative pre-trained features, reducing redundancy and controlling costs. RRS adaptively selects discriminative residuals for anomaly identification. RealNet surpasses existing methods on benchmark datasets and introduces the SIA for anomaly synthesis. Evaluation includes FID metrics and comparisons with other methods like RDR and RLPR.
RealNet outperforms the current state-of-the-art Image AU-ROC and Pixel AUROC methods on four benchmark datasets. The RealNet framework demonstrates significant improvements in both Image AU-ROC and Pixel AUROC compared to the current state-of-the-art methods. RealNet achieves substantial performance improvement compared to previous reconstruction-based methods. The results show that RealNet performs better than alternative methods such as PatchCore, SimpleNet, and FastFlow. The evaluation of the quality of anomaly images generated by RealNet using FID (Frechet Inception Distance) shows that the synthetic anomaly images are close to the distribution of real anomaly images.
In conclusion, RealNet is a cutting-edge framework for self-supervised anomaly detection comprising three key elements: SDAS, AFS, and RRS. Together, these components empower RealNet to leverage large-scale pre-trained models effectively for anomaly detection while ensuring computational efficiency. It offers a versatile platform for future anomaly detection research, particularly focusing on pre-trained feature reconstruction techniques. Extensive experiments demonstrate RealNet’s capability to tackle various real-world anomaly detection scenarios with proficiency and effectiveness.
Check out the Paper and Github. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter. Join our Telegram Channel, Discord Channel, and LinkedIn Group.
If you like our work, you will love our newsletter..
Don’t Forget to join our 38k+ ML SubReddit
Sana Hassan, a consulting intern at Marktechpost and dual-degree student at IIT Madras, is passionate about applying technology and AI to address real-world challenges. With a keen interest in solving practical problems, he brings a fresh perspective to the intersection of AI and real-life solutions.
