Improving Industrial Inspection Efficiency by Using Model Uncertainty

Abstract

Industrial inspection is an important task to guarantee the quality of products. However, even after the introduction of an anomaly detection system, visual inspection is still performed by humans. In this study, we propose a method to reduce the number of human visual inspections images and to detect mainly misclassified images. Then, we quantify the reliability of anomaly detection by using model uncertainty. In experiments, we show that the proposed method has a high detection rate of misclassified images in the number of human visual inspections.