Exploiting Training Stage Reconstruction Instability for Image Anomaly Detection

Abstract

Image anomaly detection is an important and practically significant research area. Methods such as reconstruction and feature extraction are mainly utilized. For reconstruction methods, purely normal input images should be provided, which is close to supervised learning. For feature extraction methods, unsupervised learning ones without a designated feature learning orientation perform much worse than supervised learning ones. In this paper, the goal is to achieve anomaly detection based on unsupervised learning. Generative Adversarial Network (GAN) is used as the approach, similar to unsupervised learning feature extraction methods, but it focuses more on the feature variation rate rather than using the feature outputs for further classification. The results show an innovative phenomenon in the GAN training process and achieve clear anomaly detection in unsupervised learning. The feature variations reflect the learning state of the model. The feature variations of similar images will converge in the certain learning state, while anomalies will appear as outliers. This phenomenon reveals how the model is learning, particularly about the details. The learning process of the model is intermittent and tries to learn the main features before the details. A more difficult task makes the fluctuations more significant.