A graph model-based multiscale feature fitting method for unsupervised anomaly detection. (June 2023)

• Record Type: Journal Article
• Authors: Zhang, Fanghui; Kan, Shichao; Zhang, Damin; Cen, Yigang; Zhang, Linna; Mladenovic, Vladimir
• Is Part Of: Pattern recognition. Volume 138(2023)
• Publication Date: 2023-06
• Subjects: Anomaly detection -- Unsupervised learning -- Graph model -- Feature fitting representation; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2023.109373
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: A graph model-based multiscale feature fitting method is proposed for unsupervised anomaly detection to detect and localize anomalies. The graph model between the query set and the gallery set is established according to the K nearest neighbor method. The feature fitting representation of each vertex is calculated based on its KNN features according to the message flow in the graph model. The idea of weighted multiscale anomaly map matching is proposed to obtain the anomaly map of each test image. Abstract: Anomaly detection and localization without prior knowledge is a challenging problem in industrial manufacturing due to the complexity and variety of anomaly types. Most of the existing methods have achieved considerable anomaly detection performance based on the distance between normal features and abnormal features. However, when the defect area is hard to distinguish from the background or the defect area is small, the distance between normal and abnormal features will be too close to detect anomaly areas. In addition, existing methods do not consider the influences of features in different layers with different anomaly sizes. In this paper, a graph model-based multiscale feature fitting method is proposed for unsupervised anomaly detection. Specifically, we build a graph model based on the K nearest neighbors of an anchor image. The feature fitting and anomaly scores of the anchor images in the graph vertices are calculated next. Finally, a weighted multiscale anomaly map matching method is proposed to detect and locate the anomaly regions of test images. Compared with the state-of-the-art methods, our proposed method achieves competitive improvement in anomaly detection and localization on the MVTec AD dataset, the two KolektorSDD datasets, and the mSTC dataset.