Zhixing Li

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.

In the field of intelligent manufacturing, image anomaly detection plays a pivotal role in automated product quality inspection. Most existing anomaly detection methods are adept at capturing local features of images, achieving high detection accuracy for structural anomalies such as cracks and scratches. However, logical anomalies typically appear normal within local regions of an image and are difficult to represent well by the anomaly score map, requiring the model to possess the capability to extract global context features. To address this challenge while balancing the detection of both structural and logical anomalies, this paper proposes a lightweight anomaly detection framework built upon EfficientAD. This framework integrates the reconstruction difference constraint (RDC) and a logical anomaly detection module. Specifically, the original EfficientAD relies on the coarse-grained reconstruction difference between the student and the autoencoder to detect logical anomalies; but, false detection may be caused by the local fine-grained reconstruction difference between the two models. RDC can promote the consistency of the fine-grained reconstruction between the student and the autoencoder, thereby effectively alleviating this problem. Furthermore, in order to detect anomalies that are difficult to represent by feature maps more effectively, the proposed logical anomaly detection module extracts and aggregates the context features of the image, and combines the feature-based method to calculate the overall anomaly score. Extensive experiments demonstrate our method’s significant improvement in logical anomaly detection, achieving 94.2 AU-ROC on MVTec LOCO, while maintaining strong structural anomaly detection performance at 98.4 AU-ROC on MVTec AD. Compared to the baseline, like EfficientAD, our framework achieves a state-of-the-art balance between both anomaly types.