Towards automatic crack segmentation in 3d concrete images

Concrete is one of the most commonly used construction materials. A deeper insight into its mechanical properties, in particular cracking behaviour, can be gained from stress tests. Computed tomography captures the microstructure of building materials, including crack initiation and propagation in a fully three-dimensional manner. However, the complex microstructure of concrete renders crack segmentation a very challenging task. Both, the validation of segmentation methods and the training of machine learning approaches, are hindered by the lack of reliable ground truth segmentations for real data sets. To overcome this problem, a novel procedure for generating pairs of semi-synthetic images and ground truth was introduced by the authors in a previous study. Using this semi-synthetic data, Hessian-based percolation and 3d U-net were identified as the most promising of eight approaches for crack segmentation. Here, we discuss adaptions of the methods that allow for a handling of additional features observed in real computed tomography data of concrete, in particular local variations in crack thickness.