Semi-Supervised Anomaly Detection Based on Deep Generative Models with Transformer

In this work, we propose a novel semi-supervised anomaly detection approach based on deep generative models with Transformers for identifying unusual (abnormal) images from normal ones. Our approach is based on the combination of autoencoder (AE) and generative adversarial networks (GAN). Similar to the vanilla GAN, our model is mainly composed of the generator and discriminator. The generator adopts an encoder-decoderencoder structure to extract meaningful latent representations, in which each encoder is constructed by a Transformer whereas the decoder is realized through the transposed convolution. The discriminator, which is built upon another Transformer, is used to distinguish whether the given image comes from the generator or the training set, while optimizing the encoder in the generator for better latent representations through adversarial training. The distribution of the normal data can be learned by minimizing the gap between the original image space and the latent image space during the training process. The abnormal images are detected if their distributions are different from the learned normal distributions. The merits of the proposed anomaly detection approach are demonstrated by comparing it with other generative anomaly detection approaches through experiments on three benchmark image data sets.