DRAEM -- A discriminatively trained reconstruction embedding for surface anomaly detection

Visual surface anomaly detection aims to detect local image regions that deviate from normal appearance. Recent surface anomaly detection rely on generative models to accurately reconstruct the normal areas to fail on anomalies. These methods are trained only on anomaly-free, and often require hand-crafted post-processing steps to localize the, which prohibits optimizing the feature extraction for maximal capability. In addition to reconstructive approach, we cast surface detection primarily as a discriminative problem and propose a trained reconstruction anomaly embedding model (DRAEM). The method learns a joint representation of an anomalous image and its-free reconstruction, while simultaneously learning a decision boundary normal and anomalous examples. The method enables direct anomaly without the need for additional complicated post-processing of the output and can be trained using simple and general anomaly simulations. the challenging MVTec anomaly detection dataset, DRAEM outperforms the state-of-the-art unsupervised methods by a large margin and even detection performance close to the fully-supervised methods on the used DAGM surface-defect detection dataset, while substantially them in localization accuracy.