Patch-Wise Augmentation for Anomaly Detection and Localization

Existing visual anomaly detection (AD) methods typically train reconstruction models by relying on normal images, while anomalous regions would not be well-recovered and hence can be localized with complicated post-processing steps during inference. In this paper, we formulate unsupervised AD as a supervised object detection task. To create supervision signals, we build a patch-wise data augmentation strategy called PatchAnomaly, to synthesize anomaly-like images based on self-supervised learning. Then, we propose a reconstruction-detection model to directly localize anomalous regions under supervision signals derived from PatchAnomaly. Experiments on the MVTecAD and BTAD datasets demonstrate competitive performance, achieving image-level AUROC scores of 98.4% and 95.5% respectively.