Multi-class Unsupervised Anomaly Detection with Directional Feature Perturbation

Abstract Unsupervised anomaly detection is a key 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 multi-class scenarios. Directly applying single-class methods to multi-class scenarios leads to significant performance degradation, as the model learns diverse feature information during training phase, which further amplifies the "identity shortcut" issue. To address this challenge, this paper proposes a multi-class anomaly detection model with directional feature perturbation (MAD-DFP). We regard the "identity shortcut" issue as the over-generalization of the reconstruction model to anomaly features and propose targeted strategies from the perspective of feature perturbation. Specifically, MAD-DFP locates key feature regions under the guidance of attention and perturbs features precisely, thereby suppressing the model's over-generalization problem. Meanwhile, an anomaly maps fusion module is proposed to adaptively fuse multi-level feature maps, avoiding information loss caused by simple averaging. Extensive experiments on MVTec-AD, VisA, BTAD, and MPDD show that MAD-DFP achieves optimal image-level and pixel-level performance in multi-class scenarios. Notably, the model exhibits a slight performance gap between single-class and multi-class settings, demonstrating strong application potential.