OLED: One-Class Learned Encoder-Decoder Network with Adversarial Context Masking for Novelty Detection

Novelty detection is the task of recognizing samples that do not belong to\nthe distribution of the target class. During training, the novelty class is\nabsent, preventing the use of traditional classification approaches. Deep\nautoencoders have been widely used as a base of many unsupervised novelty\ndetection methods. In particular, context autoencoders have been successful in\nthe novelty detection task because of the more effective representations they\nlearn by reconstructing original images from randomly masked images. However, a\nsignificant drawback of context autoencoders is that random masking fails to\nconsistently cover important structures of the input image, leading to\nsuboptimal representations - especially for the novelty detection task. In this\npaper, to optimize input masking, we have designed a framework consisting of\ntwo competing networks, a Mask Module and a Reconstructor. The Mask Module is a\nconvolutional autoencoder that learns to generate optimal masks that cover the\nmost important parts of images. Alternatively, the Reconstructor is a\nconvolutional encoder-decoder that aims to reconstruct unperturbed images from\nmasked images. The networks are trained in an adversarial manner in which the\nMask Module generates masks that are applied to images given to the\nReconstructor. In this way, the Mask Module seeks to maximize the\nreconstruction error that the Reconstructor is minimizing. When applied to\nnovelty detection, the proposed approach learns semantically richer\nrepresentations compared to context autoencoders and enhances novelty detection\nat test time through more optimal masking. Novelty detection experiments on the\nMNIST and CIFAR-10 image datasets demonstrate the proposed approach's\nsuperiority over cutting-edge methods. In a further experiment on the UCSD\nvideo dataset for novelty detection, the proposed approach achieves\nstate-of-the-art results.\n