Non-destructive testing (NDT) method such as computer tomography (CT) are more and more used as imaging techniques to analyze the inner structure and to detect flaws in concrete specimen. In this project algorithms have been developed to find cracks automatically in CT images of concrete. The goal is to quantify damage processes in cementious material in space and time. Three different crack detection methods have been evaluated. The statistical evaluation of detected cracks and the tracking of single cracks over several frost-thaw cycles are future project goals.


Analyzing damages at concrete structures due to physical, chemical and mechanical exposures need the application of innovative non-destructive testing (NDT) methods to trace spatial changes of microstructures. The images are obtained by 3D computer tomography with a very high resolution. This leads to data sets of more than 6 GB per concrete specimen.


  • Development of algorithms to find cracks automatically in noisy CT images of concrete
  • Quantification of damage processes in cementitious material in space and time


Three different crack detection methods for the analysis of computed tomograms of various cementitious building materials have been evaluated.

Known Methods:

The following methods have been evaluated:

  • Template Matching, which finds points where the image I is similar to a template T representing a crack
  • Filtered Hessian Eigenvalues, where the eigenvalues of the Hessian matrix for every voxel are filtered to find sheet-like structures
  • Percolation, an originally 2D algorithm based on the  physical model of liquid permeation has been extended to 3D, while the 2D circularity test to check the resulting region couldn’t  be adopted to 3D easily

New Method: Hessian-driven Percolation

  • Result of the sheet-filter computed from the Hessian eigenvalues is used to extend the percolation algorithm to 3D
  • Shape of percolated regions is checked by the coincidence with the result of the sheet filter (instead by the circularity as in the 2D case)
  • Speed is improved by starting the percolation process only at points detected by the sheet filter


3D visualization of a crack. Dataset courtesy of BAM
Template matching: Even thin cracks are detected; bending and branching parts of cracks are missed; lots of parameters; slow
Hessian Eigenvalues: Has also problems with detecting bending and branching cracks; often edges are falsely regarded as cracks; fewer parameters compared to template matching
Percolation as a stand-alone method lacks a proper 3D test if the result resembles a crack and the computation takes very long
Hessian-driven Percolation: No problems with bending and branching cracks; may bleed out, which leeds to a higher false positive rate; detected cracks are often jaggied, no detection of thin cracks; faster than template matching

Cracks are represented as a labelfield independent of the choosen detection method.

One should also take memory consumption into account when assessing the methods. Here, template matching outreaches the Hessian-based methods since the latter may produce results which need up to 50 times the size of the input dataset (Eigenvalues and Eigenvectors). Whereas the memory consumption of template matching is increased by a factor of 4. Due to the lack of reference samples and standardized image quality evaluation procedures, the results have been compared with manually segmented reference data sets. A specific question is how automatic crack detection can be used for the quantitative characterization of damage processes, such as crack length and volume. The crack detection methods have been integrated to the scientific visualization system ZIBAmira that allows displaying the tomography images as well as presenting the results.     


  • Hessian-driven percolation may serve as a reliable crack detection method for bending and branching cracks if computation speed is the main criterion
  • Template matching achieves the better results if accuracy and further crack investigations are the main goal

This is in contrast to previously given statements, but new results obtained from new datasets and computations convinced us tho change our opinion.


To assess the concrete quality not only the number and size of cracks is important but also the material the cracks run through. This is called the embedding of the cracks. The images below sketch how this is done by evaluating  the crack profiles.

Top left: Labelled crack. Top right: Crack labelled according to the embedding. Bottom left: probelines orthogonal to crack voxels. Bottom middle and right: typical crack probelines (on aggregate border, inside aggregate)


  • Tracing of single cracks over time
  • Using the result of the crack detection for further analysis, e.g. crack width or crack orientation
  • Representation of cracks as point clouds and polygonal structures