SLSG: Industrial image anomaly detection with improved feature embeddings and one-class classification

Minghui Yang, Jing Liu, Zhiwei Yang, Zhaoyang Wu

Abstract: Highlights•Generative pre-training network can better do feature embedding of normal pattern.•Anomaly simulation can assist in the learning of one-class classification.•Self-attention-based GCN can enhance the detection of model for logical anomaly.•SLSG achieves SOTA performance and faster speed in detecting comprehensive anomaly.