Cascade RDN: Towards Accurate Localization in Industrial Visual Anomaly Detection With Structural Anomaly Generation

Abstract

Unsupervised visual anomaly detection uses only anomaly-free images to detect anomalous patterns, whose recent methods mainly focus on the anomaly classification sub-task but neglect to localize anomalies accurately. Existing reconstruction-based and representation-based methods yield anomaly score maps that often predict erroneous responses. The method jointly trained with the discriminative network is not suitable for anomalies other than texture types. This letter propose a novel cascade reconstruction-discriminant network (Cascade RDN), which adopts a cascade structure to obtain representative discriminative reconstruction embedding. The bi-direction channel attention module is designed to enable the two discriminative sub-networks to boost each other. Moreover, a general structural anomaly generation is presented to complement the existing textured anomaly generation to cover all types of surface anomalies. The proposed method outperforms previous anomaly localization methods by 7%-10% in AP on two challenging benchmarks MVTec AD and BTAD. Meanwhile, it achieves state-of-the-art anomaly classification scores on BTAD.