Adversarially learned one-class novelty detection with confidence estimation

Abstract

Given samples from a particular normal class, image novelty detection is aimed at determining whether a query sample is from the normal class. Images from this particular class are termed inliers or normal images, whereas images not belonging to the class are termed novelties. Most novelty detection approaches use deep neural networks. However, few of them are end-to-end, and state-of-the-art neural networks tend to be overconfident in their predictions. This may result in incorrect predictions and may negatively affect novelty detection. In this paper, we propose a novel model termed adversarially learned one-class novelty detection with confidence estimation for image novelty detection. The proposed model consists of a representation and a detection module, which are adversarially trained to collaboratively learn the inlier distribution. Moreover, the model uses confidence estimation so that the detection module can more effective. The proposed model is end-to-end and does not require additional calculations such as novelty scores after training. We conduct comprehensive experiments on four publicly available datasets that are commonly used for novelty detection, and the model is compared with state-of-the-art methods to demonstrate its performance.