An Novel Anomaly Detection Method for Tiny Defects on Translucent Glass

Abstract

Translucent glass is widely used in advanced touch panel display devices. But its defects are more difficult to be found because of variances of the transparency and shape. In this paper, we propose an novel CFLOW-based unsupervised anomaly detection method. We improved the multi-scale aggregation strategy by introducing deeper layer and shallow layer likelihood differences into the final anomaly map. This method reduces the defect detection area on translucent glass from hundreds of pixels to less than 10 pixels. Finally, our method improves the average area under the receiver operating characteristic curve to 96.74% at the image level and 98.96% at the pixel level, which is satisfactory for industrial applications.