Anomaly detection for industrial quality assurance: A comparative evaluation of unsupervised deep learning models

Abstract

Across many industries, visual quality assurance has transitioned from a manual, labor-intensive, and error-prone task to a fully automated and precise assessment of industrial quality. This transition has been made possible due to advances in machine learning in general, and supervised learning in particular. However, the majority of supervised learning approaches only allow to identify pre-defined categories, such as certain error types on manufactured objects. New, unseen error types are unlikely to be detected by supervised models. As a remedy, this work studies unsupervised models based on deep neural networks which are not limited to a fixed set of categories but can generally assess the overall quality of objects. More specifically, we use a quality inspection case from a European car manufacturer and assess the detection performance of three unsupervised models (i.e., Skip-GANomaly, PaDiM, PatchCore). Based on an in-depth evaluation study, we demonstrate that reliable results can be achieved with fully unsupervised approaches that are even competitive with those of a supervised counterpart.