DualFlow: Dual-Branch Flow for Unsupervised Anomaly Detection and Localization

Industrial production processes present challenges in collecting and annotating defect samples, primarily due to cost constraints and the inability to comprehensively cover all defect categories. To address this issue, recent research has focused on modeling normal sample features. This article proposes DualFlow, an unsupervised anomaly detection and localization algorithm based on normalizing flow. DualFlow uses a dual-branch architecture to effectively balance the detection and localization capabilities of existing algorithms. To handle features of different scales, we propose a compact and efficient module called the gated multiscale feature fusion (GMFF) module. In addition, DualFlow incorporates a variance stability loss to exploit the inherent stability of normal sample features, resulting in a significant reduction in false-positive instances. DualFlow is designed for end-to-end training and inference, greatly enhancing defect detection and localization capabilities. DualFlow achieved 99.33% image-level area under the receiver operating characteristic (AUROC) and 98.05% pixel-level AUROC on the MVTec AD dataset. Furthermore, DualFlow exhibits robust generalization performance, as confirmed by its evaluation on the more challenging MVTec LOCO AD, BTAD, VisA, and KolektorSDD2 datasets.