Defect detection in CT scans of cast aluminum parts

• Verfasst von: Fuchs, Patrick [VerfasserIn]
• Verfasst von: Kröger, Thorben [VerfasserIn]
• Verfasst von: Garbe, Christoph S. [VerfasserIn]
• Titel: Defect detection in CT scans of cast aluminum parts
• Titelzusatz: a machine vision perspective
• Verf.angabe: Patrick Fuchs, Thorben Kröger, Christoph S. Garbe
• E-Jahr: 2021
• Jahr: 29 April 2021
• Umfang: 12 S.
• Teil: volume:453
• Teil: year:2021
• Teil: pages:85-96
• Teil: extent:12
• Fussnoten: Gesehen am 14.07.2021
• Titel Quelle: Enthalten in: Neurocomputing
• Ort Quelle: Amsterdam : Elsevier, 1989
• Jahr Quelle: 2021
• Band/Heft Quelle: 453(2021), Seite 85-96
• ISSN Quelle: 1872-8286
• DOI: doi:10.1016/j.neucom.2021.04.094
• Datenträger: Online-Ressource
• Sprache: eng
• Sach-SW: Computed tomography
• Sach-SW: Deep learning
• Sach-SW: Defect detection
• Sach-SW: Self-supervision
• Sach-SW: Semantic segmentation
• Sach-SW: Simulated training data
• K10plus-PPN: 1762909502

Abstract

One of the many applications of X-ray computed tomography (CT) in industry is the detection of pores, cavities and other flaws in cast metal parts. Because of its improvement on part safety and saving of expenses, CT inspection is moving from a random sample inspection towards a full in-line inspection. With the increasing amount of produced data, however, comes the need for an automated processing. Due to tight time constraints the resulting CT scans are very artifact afflicted, which impedes automated inspection. In recent years, deep learning methods—convolutional neural networks in particular—have been used with great success to tackle even complex segmentation tasks in cluttered scenes. As we show, these methods are also applicable to the domain of industrial CT data: they are able to cope with noise, beam hardening, scatter and other artifacts which we encounter here. However, these methods need a vast amount of precisely labeled training data to work properly. Gathering the necessary data is not only cumbersome due to the need of annotating three-dimensional data but also expensive as it requires the knowledge of domain experts. Therefore, we present a new approach: We train our models on realistically simulated CT data only. Here, a precise per-voxel ground truth can simply be computed. In order to show that the simulated data is sufficient to train a segmentation network, we turn to its prediction performance on real CT data. We compare the prediction performance of traditional algorithms as well as the trained segmentation network on simulated and real validation data and demonstrate that they behave similarly. The ground truth for the real validation data is hand-labeled using high-quality CT scans, while the actual validation set consists of CT scans of lower quality of the exact same parts. For a comprehensive evaluation, we evaluate the probability of detection as well as the intersection over union. The first tells us how likely a flaw of given size can be found with a given confidence, which is of special interest to domain experts. The latter gives us a per-voxel information of how precise the overall segmentation is. Moreover, our synthetic data enables us to examine the influence of different artifact types on the detection rate. Besides these quantitative analyses we show some qualitative results of real-world applications. To the best of our knowledge, we describe the first approach for defect detection in three-dimensional CT data, which is solely trained with simulated data.