The classification of microstructures in flake graphite cast iron (GJL) according to DIN EN ISO 945-1 is a challenging task due to the presence of mixed microstructures and the smooth transitions between different graphite arrangement classes. Traditional manual evaluations are subjective and often exhibit inconsistencies between different experts and even within repeated assessments by the same individual. To address this challenge, we propose an object-based semantic segmentation approach utilizing a U-Net neural network.

Our methodology involves the generation of a synthetic training dataset by extracting individual microstructure objects from real micrographs, assigning them to predefined classes, and assembling them into synthetic images with corresponding segmentation masks. A U-Net model is trained on this dataset to enable objective, reproducible, and automated classification of GJL micrographs. The evaluation on synthetic test data demonstrates high segmentation accuracy. However, challenges remain in generalizing the approach to real-world micrographs due to variations in sample preparation, image quality, and scaling factors.

Future research will focus on enhancing classification accuracy by expanding the dataset, incorporating three-dimensional structures, and integrating global texture features to improve segmentation performance on real micrographs.