AI based particle separation in computed tomographic images of aggregates

• Applicant: Professor Dr.-Ing. Wolfgang Breit
• Subject Area: Construction Material Sciences, Chemistry, Building Physics ; Methods in Artificial Intelligence and Machine Learning
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 558673683

Aggregates are an important component of concrete and numerous other building materials. The choice of aggregate, its composition, size distribution and shape also have a significant influence on the properties of the resulting products. In order to take these influences into account, it is necessary to characterize the aggregates. Currently, this normative characterization is usually carried out using manual methods, such as the sieving method for determining the particle size distribution and various methods for determining the particle shape. Other methods, including automated methods, have not yet been generally accepted. An alternative method that is suitable for characterizing aggregates is computed tomography (CT). It has already been proven that, using CT and 3D image analysis, morphological characteristics of the aggregate can be obtained, that have a direct influence on the resulting fresh and hardened material properties of mineral products. However, characterization via CT and analysis of the resulting 3D images is only practicable if the grain shape characteristics of the individual grains can also be correctly determined within a sample with many aggregates as bulk material. To do this, it is necessary to separate the individual particles using image analysis. This is however difficult due to the widely varying shapes and sizes. Classical particle separation using the watershed transformation on the inverted distance image suffers from over-segmentation. The adaptive h-extrema transformation can reduce this considerably, but its correct application and parameterization requires expert knowledge. In addition, there are still too many incorrect segmentations that require time-consuming manual post-processing. As part of preliminary work, random forests were therefore trained to automatically detect over-segmentation based on the characteristics of triples consisting of two grains/fragments and the watershed separating them. This has already enabled 27 data sets of different aggregates, including fine sands and recycled materials, to be correctly segmented. Currently, however, this approach still requires a combination of several software tools and the selection of a trained random forest. The aim of the planned project is therefore to develop a prototype software solution and a corresponding process specification that will make the process potentially usable for industrial practice.

• DFG Programme: Research Grants (Transfer Project)
• Application Partner: Basalt-Actien-Gesellschaft (BAG); Südwestdeutsche Hartsteinwerke
• Cooperation Partner: Dr. Katja Schladitz