Robust Anomaly Detection and Localization via Simulated Anomalies

Anomaly detection refers to identifying abnormal images and localizing anomalous regions. Reconstruction-based anomaly detection is a commonly used method; however, traditional reconstruction-based methods perform poorly as deep models generalize successfully enough that even anomalous regions can be well-restored. In this paper, we propose a new method to address the single pseudo-anomaly type and high false positive detection of the existing methods. Specifically, we design a novel pseudo-anomaly simulation module that can generate several types of anomalies on normal images. Furthermore, we propose an effective reconstruction network to improve the robustness of the model against distractors. Finally, we employ a segmentation network to localize anomalous regions. This simple but effective method can detect various anomalies in the real world, even those that are subtle and rare. Extensive experiments on the MVTec anomaly detection dataset demonstrate the effectiveness and superiority of the proposed method, yielding an AUROC score of 98.2% in image-level anomaly detection and 97.8% in pixel-level anomaly localization.