A Transformer-Based GAN for Anomaly Detection

AbstractAnomaly detection is the task of detecting outliers from normal data. Numerous methods have been proposed to address this problem, including recent based on generative adversarial network (GAN). However, these are limited in capturing long-range information data due receptive field obtained by convolution operation. The crucial for producing distinctive representation belonging different classes, while local important distinguishing abnormal data, if they belong same class. In paper, we propose a novel Transformer-based architecture anomaly which has advantages extracting features with global representing classes as well details useful anomalies. our design, introduce self-attention mechanism into generator GAN extract semantic information, and also modify skip-connection capture multi-scale input experiments CIFAR10 STL10 show that method provides better performance compared state-of-the-art CNN-based methods. Experiments performed MVTecAD LBOT datasets offers results, outperforming baseline SAGAN over 3% terms AUC metric.KeywordsAnomaly detectionTransformerGenerative advertise