Anomaly detection in images with shared autoencoders

Abstract

Anomaly detection is a classical problem in computer vision, namely the determination of the normal from the abnormal when datasets are highly biased toward one class (normal) due to the insufficient sample size of the other class (abnormal). We introduce a novel model that utilizes two decoders to share two encoders, respectively, forming two sets of network structures of encoder-decoder-encoder called EDE, which are used to map image distributions to predefined latent distributions and vice versa. In addition, we propose an innovative two-stage training mode. The first stage is roughly the same as the traditional autoencoder (AE) training, using the reconstruction loss of images and latent vectors for training. The second stage uses the idea of generative confrontation to send one of the two groups of reconstructed vectors into another EDE structure to generate fake images and latent vectors. This EDE structure needs to achieve two goals to distinguish the source of the data: the first is to maximize the difference between the fake image and the real image; the second is to maximize the difference between the fake latent vector and the reconstructed vector. Another EDE structure has the opposite goal. This network structure combined with special training methods not only well avoids the shortcomings of generative adversarial networks (GANs) and AEs, but also achieves state-of-the-art performance evaluated on several publicly available image datasets.