A Unified Model for Multi-class Anomaly Detection

Despite the rapid advance of unsupervised anomaly detection, existing methodsrequire to train separate models for different objects. In this work, wepresent UniAD that accomplishes anomaly detection for multiple classes with aunified framework. Under such a challenging setting, popular reconstructionnetworks may fall into an "identical shortcut", where both normal and anomaloussamples can be well recovered, and hence fail to spot outliers. To tackle thisobstacle, we make three improvements. First, we revisit the formulations offully-connected layer, convolutional layer, as well as attention layer, andconfirm the important role of query embedding (i.e., within attention layer) inpreventing the network from learning the shortcut. We therefore come up with alayer-wise query decoder to help model the multi-class distribution. Second, weemploy a neighbor masked attention module to further avoid the information leakfrom the input feature to the reconstructed output feature. Third, we propose afeature jittering strategy that urges the model to recover the correct messageeven with noisy inputs. We evaluate our algorithm on MVTec-AD and CIFAR-10datasets, where we surpass the state-of-the-art alternatives by a sufficientlylarge margin. For example, when learning a unified model for 15 categories inMVTec-AD, we surpass the second competitor on the tasks of both anomalydetection (from 88.1% to 96.5%) and anomaly localization (from 89.5% to 96.8%).Code is available at https://github.com/zhiyuanyou/UniAD.