SFB/TRR 280 (Design Strategies for Material-Minimised Carbon Reinforced Concrete Structures) - Image Analysis Methods in Micro-Tomography

Resarcher: Frank Liebold, Franz Wagner, Hannes Sardemann
Former: Ferdinand Maiwald, Ariel Rodriguez

This project is part of a Collaborative Research Centre SFB/TRR 280 (Design Strategies for Material-Minimised Carbon Reinforced Concrete Structures—Principles of a New Approach to Construction, subprojects D03 and S) and it is funded by the German Research Council (Deutsche Forschungsgemeinschaft, DFG). Several parts of the project will be presented in the following points.

The Project

In close cooperation with the other 21 subprojects at the Dresden and Aachen, new structural shape and design strategies are being developed and implemented. However, the carbon reinforced concrete structures must be comprehensively investigated in order to ensure their reliability. The subprojects D03 and S focus on the recording and 3D reconstruction of such samples by a µ-tomograph up to the structural analysis of the resulting data.

Computertomography

A tomography device consists of an X-ray source and an image plane, which can be seen as a camera. During the scan, a large number of projections are aquired while the object rotates.

From these images, a volumetric reconstruction of the object is obtained, which is then represented by a grayscale voxel space, wherein the elements contained in the specimen are visible.

Deep Learning based 3D Structural Analysis

The resulting reconstructions can then be analyzed in terms of their components, properties, and deformations. Quantitative structural analysis is performed using 3D Convolutional Neural Networks (CNN), as a process based on simpler algorithms is not able to confidently segment the smallest constituents.

First, the neural networks are trained in a supervised manner. This means that there must already be a segmented volume on which the network learns which structures it should segment.

However, the production of the training data is extremely time-consuming, which is why an unsupervised (without the influence of humans) learning process is developed. This eliminates the need for future processors to manually produce the training data. The goal is achieved by a 3D Generative Adversarial Network (GAN), which generates deceptively real training data.

3D Displacement Fields between Deformation States

There exist at least two volumes of different load steps of a probe.

Cross-sections of the reference and the deformed state:

In the reference volume a grid of point is defined. In order to obtain the displacement field for these points, 3D least squares matching is applied. It is a gradient based subpixel precise matching method.

Deformation Analysis

The appearance of cracks leads to discontinuities in displacement field. The following deformation vectors are determined:

Publications 2022

Liebold, F., Maas, H.-G. (2022):
Computational Optimization of the 3D Least-Squares Matching Algorithm by Direct Calculation of Normal Equations. Tomography 8(2), 760-777, https://doi.org/10.3390/tomography8020063.

Liebold, F., Maas, H.-G. (2022):
3D-Deformationsanalyse und Rissdetektion in multitemporalen Voxeldaten von Röntgentomographen. Dreiländertagung der DGPF, der OVG und der SGPF in Dresden – Publikationen der DGPF,  Band 30, 105–116, https://doi.org/10.24407/KXP:1796026123.

Publications 2021

• Liebold, F., Lorenzoni, R., Curosu, I., Léonard, F., Mechtcherine, V., Paciornik, S., Maas, H.-G. (2021):
  3D Least Squares Matching Applied to Micro-Tomography Data. ISPRS - International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XLIII-B2-2021, 533–539, https://doi.org/10.5194/isprs-archives-XLIII-B2-2021-533-2021.