Adaptive Unsupervised Anomaly Detection in Variable Environment by Online Expectation Maximization

Automatic anomaly detection (AD) in a series of images of industrial parts is a key component of industrial production and an exemplary problem for machine learning. Since it can only realistically function with minimal supervision, unsupervised methods dominate the field. Their principle is that the ``normal aspect"" of objects is learned from recently observed samples, so that anomalies can be detected as outliers. In this paper, we start by reviewing recent AD methods and their performance-based ranking on recent benchmark datasets. The recent progress of such methods is such that they learn from a few hundred normal samples only. However, we argue that the current method evaluation based on static datasets is limited and biased. Indeed, a main feature of industrial production is that the aspect of objects evolves over time, due to changes in production and acquisition conditions, thus leading to significant probability distribution shifts. By introducing artificial but realistic deviations into the classic MVTec benchmark we show that the smallest deviation is sufficient to make these stationary models collapse. We argue that some of these models, especially the stochastic ones, can be easily adapted to cope with distribution shifts. The Global-to-Local Anomaly Detector (GLAD) is such an example of a method that uses Gaussian Mixture Models to model the distribution of regular objects. Using the stochastic approximation of expectation maximization, we design Online-GLAD, an improved GLAD that can update and adapt online. In the experiments, we show that Online-GLAD is able to maintain good performance even in the presence of multiple progressive deviations, and with constant complexity compatible with real-time implementation.