Cracks in concrete

TL;DR

This study tackles the challenge of segmenting cracks in 3D CT images of concrete, where cracks are thin, dark, and embedded in a highly heterogeneous matrix. It presents a semi-synthetic data generation pipeline using crack models (Fractional Brownian motion, Voronoi-based surfaces) and adaptive thickness to train 3D CNNs, including a 3D U-Net and multiscale variants, for crack segmentation. To address real-world variability, the authors introduce multiscale image pyramids and a scale-invariant Riesz network, showing robustness across synthetic multiscale cracks and comparing favorably with standard 3D U-Nets. They further extend the approach to fiber-reinforced concrete, demonstrating the need for background-aware training data. The work highlights the potential of scale-invariant, data-efficient architectures and synthetic data in enabling reliable crack detection, with ongoing challenges in generalization to new concrete formulations and practical calibration.

Abstract

Finding and properly segmenting cracks in images of concrete is a challenging task. Cracks are thin and rough and being air filled do yield a very weak contrast in 3D images obtained by computed tomography. Enhancing and segmenting dark lower-dimensional structures is already demanding. The heterogeneous concrete matrix and the size of the images further increase the complexity. ML methods have proven to solve difficult segmentation problems when trained on enough and well annotated data. However, so far, there is not much 3D image data of cracks available at all, let alone annotated. Interactive annotation is error-prone as humans can easily tell cats from dogs or roads without from roads with cars but have a hard time deciding whether a thin and dark structure seen in a 2D slice continues in the next one. Training networks by synthetic, simulated images is an elegant way out, bears however its own challenges. In this contribution, we describe how to generate semi-synthetic image data to train CNN like the well known 3D U-Net or random forests for segmenting cracks in 3D images of concrete. The thickness of real cracks varies widely, both, within one crack as well as from crack to crack in the same sample. The segmentation method should therefore be invariant with respect to scale changes. We introduce the so-called RieszNet, designed for exactly this purpose. Finally, we discuss how to generalize the ML crack segmentation methods to other concrete types.

Figures (4)

• Figure 1: Example of a crack in concrete as they appear in optical images. Small cutout from a 2D image of a concrete panel. 600 by 600 pixels cover approximately $4.5\,$cm by $4.5\,$cm. Left: original. Right: crack as segmented by CrackNet mueller18, see Section \ref{['sec:ml-2d-crack-segmentation']}.
• Figure 2: Examples of cracks in concrete as they appear in CT images. 2D slices from the 3D images. All samples created by University of Kaiserslautern, Civil Engineering. CT scans by Fraunhofer ITWM, except for top right scanned by Fraunhofer EZRT using the setup salamon:gulliver. Left: Normal concrete. Right: Fiber (bottom steel, top polymer) reinforced concrete. Section images contain $1\,200\times1\,200$ pixels covering squares of edge lengths between 2.7 and 4.3 cm.
• Figure 3: Examples of simulated cracks in $256\times256\times256$ voxel images. Left: two fractional Brownian surfaces, both widened to constant thickness 3 pixels and with Hurst index $0.97$. Right: a minimal surface from a spatial Voronoi tessellation, width varying according to a Bernoulli random walk with parameter. Top: volume renderings. Bottom: 2D slices of the 3D images of the cracks superimposed on CT images of high performance concrete, scanned with a voxel size of $23.5\,$µ m. The visualized cubes and slices have therefore edge length $6\,$mm.
• Figure 4: Examples of segmentation results for fiber reinforced concrete. The 3D U-Net not trained on FRC mistakes dark regions close to the bright steel fibers as cracks (b) and does not detect the thin branches of the crack in the PPFRC (e). The fine tuned 3D U-Net performs much better although classifying some pp fibers as crack (f).