Steel is one of the most important modern materials, forming the foundation of numerous key technologies. To tailor its properties, a profound understanding of its microstructure—especially in complex bainitic–martensitic steels—is essential. However, objective and reproducible microstructure characterization remains a major challenge. In practice, light-optical microscopy (LOM) is widely used due to its simplicity, yet it suffers from limited resolution, strong dependence on etching conditions, and a lack of standardized classification criteria. Expert assessments are often subjective and inconsistent, and no broad consensus exists within the field.

A particular difficulty in analyzing quenched martensitic–bainitic microstructures is the creation of pixel-wise annotations required for conventional semantic segmentation approaches. In such highly complex systems, this is often infeasible due to blurred phase boundaries and ambiguous regions. The presented approach addresses this by employing partially annotated ground truth masks based on correlative microscopy (LOM, SEM, and EBSD), combined with tailored loss functions that consider only the confidently labeled regions during training. This enables a deep learning model to be trained on objective, crystallographically informed reference data, ultimately allowing for efficient and reproducible phase segmentation in purely light-optical micrographs. The result is a scalable and practical solution for the quantitative evaluation of complex steel microstructures based on simple LOM images.