Simulation of microstructures and machine learning

TL;DR

The paper addresses data scarcity in ML for imaging microstructures and surface inspection by leveraging synthetic data from stochastic geometry. It analyzes three use cases: reconstruction of highly porous structures from FIB-SEM, crack segmentation in 3D concrete CT, and optical surface defect inspection. It demonstrates training ML models on synthetic data, including 3D U-net for porosity reconstruction and a novel scale-invariant RieszNet for cracks, highlighting the benefits and remaining generalization challenges. The discussion points to open questions on realism, domain gap quantification, and the need for dedicated metrics and rendering choices to reliably transfer synthetic training to real data.

Abstract

Machine learning offers attractive solutions to challenging image processing tasks. Tedious development and parametrization of algorithmic solutions can be replaced by training a convolutional neural network or a random forest with a high potential to generalize. However, machine learning methods rely on huge amounts of representative image data along with a ground truth, usually obtained by manual annotation. Thus, limited availability of training data is a critical bottleneck. We discuss two use cases: optical quality control in industrial production and segmenting crack structures in 3D images of concrete. For optical quality control, all defect types have to be trained but are typically not evenly represented in the training data. Additionally, manual annotation is costly and often inconsistent. It is nearly impossible in the second case: segmentation of crack systems in 3D images of concrete. Synthetic images, generated based on realizations of stochastic geometry models, offer an elegant way out. A wide variety of structure types can be generated. The within structure variation is naturally captured by the stochastic nature of the models and the ground truth is for free. Many new questions arise. In particular, which characteristics of the real image data have to be met to which degree of fidelity.

Figures (6)

• Figure 1: Realizations of stochastic geometry models. From left to right: Simulated SEM image of a Boolean model of cylinders, Sec. \ref{['sec:challenge-fib-sem']}. CT image of a concrete sample with impressed crack generated based on a Voronoi tessellation, Sec. \ref{['sec:challenge-cracks']}. Color coded height map of a simulated milling pattern on a metal surface, Sec. \ref{['sec:challenge-vipoi']}.
• Figure 2: Synthetic SEM images of realizations of Boolean models of spheres, cylinders, and cubes, a Cox-Boolean sphere model, and Altendorf-Jeulin's curved fiber model altendorf10:model. Images generated using Prill's method prill12:scanning.
• Figure 3: Examples of simulated cracks in $400^3$ voxel images. Minimal surfaces from spatial Voronoi tessellations generated by a Matérn cluster process (left), force-biased sphere packing (center left), a Poisson point process (center right), and a Poisson point process stretched in x and z directions (right). The crack widths vary according to a Bernoulli random walk jung22cracks with parameter $p=0.01$. Top: volume renderings. Bottom: 2D slices of the 3D images of the cracks superimposed on CT images of normal, high performance, and aerated concrete and the polypropylene fiber reinforced concrete from ict23. Voxel sizes are $23\,$µ m, $20\,$µ m, $2.8\,$µ m, and $60\,$µ m. Hence, cube and slice edge lengths vary between $1$ and $24\,$mm.
• Figure 4: Segmentation results for the high performance concrete sample from ict22.
• Figure 5: Segmentation results for the polypropylene fiber reinforced concrete sample from ict23. The RieszNet result is obtained on a downscaled version of the image and stitched. The missing part in the center of the crack is due to edge effects and the stitching.