AI-Powered Defect Segmentation in Industrial CT Data

Abstract

Non-destructive quality testing using CT plays an important role in industrial quality assurance. However, manual analysis of the large voxel data sets is not efficient. AI-powered processes have already shown that successful segmentation of defects in industrial voxel data is possible. In this paper, we show an AI-solution to detect small defects in industrial CT data. Therefore, we propose two new network architectures, PU-Net and PCU-Net based on an U-Net architecture. For this purpose, we first conducted a parameter study to determine the parameters with the greatest impact on the segmentation performance and incorporated them into the new
architecture. In addition, we improved a reference dataset by introducing a data augmentation and also improved the annotation of the real data in this dataset. The evaluation of the new architectures showed very good results.