Abstract
We present a novel method for image anomaly detection, where algorithms thatuse samples drawn from some distribution of "normal" data, aim to detectout-of-distribution (abnormal) samples. Our approach includes a combination ofencoder and generator for mapping an image distribution to a predefined latentdistribution and vice versa. It leverages Generative Adversarial Networks tolearn these data distributions and uses perceptual loss for the detection ofimage abnormality. To accomplish this goal, we introduce a new similaritymetric, which expresses the perceived similarity between images and is robustto changes in image contrast. Secondly, we introduce a novel approach for theselection of weights of a multi-objective loss function (image reconstructionand distribution mapping) in the absence of a validation dataset forhyperparameter tuning. After training, our model measures the abnormality ofthe input image as the perceptual dissimilarity between it and the closestgenerated image of the modeled data distribution. The proposed approach isextensively evaluated on several publicly available image benchmarks andachieves state-of-the-art performance.