Abstract
Anomaly detection is a classical problem in computer vision, namely thedetermination of the normal from the abnormal when datasets are highly biasedtowards one class (normal) due to the insufficient sample size of the otherclass (abnormal). While this can be addressed as a supervised learning problem,a significantly more challenging problem is that of detecting theunknown/unseen anomaly case that takes us instead into the space of aone-class, semi-supervised learning paradigm. We introduce such a novel anomalydetection model, by using a conditional generative adversarial network thatjointly learns the generation of high-dimensional image space and the inferenceof latent space. Employing encoder-decoder-encoder sub-networks in thegenerator network enables the model to map the input image to a lower dimensionvector, which is then used to reconstruct the generated output image. The useof the additional encoder network maps this generated image to its latentrepresentation. Minimizing the distance between these images and the latentvectors during training aids in learning the data distribution for the normalsamples. As a result, a larger distance metric from this learned datadistribution at inference time is indicative of an outlier from thatdistribution - an anomaly. Experimentation over several benchmark datasets,from varying domains, shows the model efficacy and superiority over previousstate-of-the-art approaches.