Just Noticeable Learning for Unsupervised Anomaly Localization and Detection

Ying Zhao

Abstract: Anomaly detection targets to detect unseen defective ap-pearances that are different from the seen non-anomalous instances. Recent unsupervised learning methods generate anomalous instances to inspire the anomaly detector to learn discriminative features. However, the generated anomalous instances are more easily noticed than real defects and may cause inconsistency between normal and anomalous. It lim-its the reconstruction of anomaly-free images and the seg-mentation of diversity anomalies. To conquer these prob-lems, we propose a Just Noticeable learning for unsupervised anomaly Localization and Detection(JNLD). The method simulates multi-scale noticeable defects to enrich generaliz-able representation. It embeds just noticeable learning in the sub-networks of reconstruction and segmentation to di-rectly localize the defects without any complex additional post-processing. With further equipment of label inconsis-tency correction, our JNLD achieves 73.8 for unseen anomaly localization average precision that surpasses the current state-of-the-art by a large margin of 5.4 percentage on the challenging dataset MVTecAD.