Multiclass unsupervised anomaly detection with directional feature perturbation

Abstract Unsupervised anomaly detection is a crucial technology in the field of intelligent manufacturing. The conventional single-class paradigm of ‘one-class-one-model’ faces challenges of surging training and deployment costs in multiclass scenarios. Directly applying single-class methods to multiclass scenarios leads to significant performance degradation, as the model learns diverse feature information during the training phase, which further amplifies the ‘identity shortcut’ issue. To address this challenge, this paper proposes a multiclass anomaly detection model with directional feature perturbation (MAD–DFP). We regard the ‘identity shortcut’ issue as an overgeneralization of the reconstruction model to anomaly features and propose targeted strategies from the perspective of feature perturbation. In particular, MAD–DFP precisely locates key feature regions under the guidance of attention and perturbs features, thereby suppressing the model’s overgeneralization problem. Meanwhile, an anomaly map fusion module is proposed to adaptively fuse multilevel feature maps, avoiding information loss caused by simple averaging. Extensive experiments on MVTec-AD, VisA, BTAD, and MPDD reveal that MAD–DFP achieves optimal image-level and pixel-level performance in multiclass scenarios. Notably, the model exhibits a slight performance gap between single-class and multiclass settings, demonstrating strong application potential.