Concrete is one of the most frequently used building materials. Since failure of concrete can have catastrophic consequences, understanding its properties is a prerequisite to wide-spread application. In particular, mechanical tests are applied to investigate the cracking behaviour of concrete. Computed tomography (CT) is used to analyze the concrete’s microstructure non-destructively.
Segmenting cracks in CT images is challenging. Crack structures may be very thin and, therefore, insufficiently resolved in the image data. Additionally, concrete is a heterogeneous composite material typically consisting of cement paste, aggregates, pores, and reinforcements. Hence, cracks have to be distinguished from these components. The immense size of representative image data (up to 2,000^2x10,000 voxels for a concrete beam) prohibits manual processing.
We compare several methods for automatic crack segmentation. To enable a fair comparison, we design a comparative study based on simulated data. Simulated cracks of varying width and shape are integrated in CT images of concrete to achieve realistic images. Using these data, machine learning methods can be trained for the segmentation. Additionally, the existence of a ground truth allows for an objective evaluation of the results. The best segmentation results are obtained by a convolutional neural network (U-net) and by Hessian-based percolation.
These two methods are applied to real CT scans. There we observe an additional challenge: The thickness of cracks varies continuously while the crack propagates. Hence, multi-scale approaches covering crack widths between 1 and 20 voxels are explored.
On large images, voxel-wise crack segmentation on the whole image is not feasible from a computational point of view. Hence, we suggest to initially roughly scan the image for regions that are likely to contain cracks. Then, crack segmentation is restricted to those regions.
Applicability, limitations, and robustness of our methods are discussed.