The basis for successful machine learning (ML) segmentation and classification is a clean, well-founded ground truth. When studying complex microstructures, the generation of a ground truth is not only time and labor-intensive but can also easily lead to disagreements between human experts based on how the microstructure appears visually to the expert's eye.

This paper presents a correlative microscopy approach that enables the automated and objective generation of annotations for segmenting or classifying complex microstructures in light optical microscope (LOM) or scanning electron microscope (SEM) images. First, the same sample area is captured by LOM or SEM and by electron backscatter diffraction (EBSD). The data from the EBSD measurements, which can be processed as image data or numerical data, are then used to generate objective masks. Both unsupervised ML approaches and simple supervised ML approaches that require only small amounts of annotated EBSD data will be used to automatically generate the annotations, which will then be used to train the ML models for segmentation and classification of LOM or SEM images. As use cases, the segmentation of multiphase steel microstructures and the classification of microstructures from quenched and tempered steels will be presented.

The automated approach based on EBSD measurements not only reduces the labor-intensive, tedious work of manually annotating images, but also increases the objectivity and reproducibility of the ground truth. It thus enables the most accurate segmentation and classification of microstructures.