A Latent Feature Autoencoder via Adversarial Training for Unsupervised Anomaly Detection

Anomaly detection is an active area of computer vision and widely applied in diverse fields. As known, it is a considerable challenge to collect abnormalities in practice. To tackle it, researchers propose many unsupervised or semi-supervised algorithms based on autoencoders or their variants. They focus on reconstruction loss between input and reconstructed samples but ignore the latent features extracted by an autoencoder. Namely, the existing algorithms tend to learn the local features of samples and have the approximate capabilities to reconstruct normal and abnormal samples. To capture the latent spatial features of anomaly detection, we present an unsupervised latent feature autoencoder via adversarial training. Particularly, we propose a weighted feature consistency loss to exploit the correlation between the corresponding layers of an encoder and decoder in the autoencoder. A feature discrimination loss is also designed to improve its ability to identify real and reconstructed samples by utilizing the latent spatial features of a discriminator. Next, we develop a discriminator that consists of two networks, i.e., a feature extraction network and a classification one. The pre-trained model can extract the input features accurately and stably, while the classification network can avoid excessive information losses and strengthen the ability to acquire deep semantics. Extensive experiments conducted on six public datasets show that the proposed method is competitive with the existing mainstream methods.