Parameters describing the microstructure play an important role in material characterization. For quantitative microstructural analysis, grain boundaries must be correctly detected as a basis for further steps. Especially for materials that are difficult to analyze (e.g. due to their nanocrystalline microstructure with limited grain boundary visibility or due to preparation artifacts), previous conventional automation approaches for this task reach their limits. In practice, therefore, grain boundary detection is often supported manually by experts. Computer vision based on machine learning represents a promising tool to exclude the subjectivity and automate this task. At Fraunhofer IKTS (Dresden, Germany), Convolutional Neural Networks are investigated for fully automated grain boundary detection. Using the example of a single-phase Al₂O₃ microstructure, a network architecture from the field of edge detection of real-life objects is trained. The necessary data set consists of microstructural images taken with two different detectors of a high resolution FE-SEM as well as the corresponding manually annotated ground truth images. In the investigations, the images of the two detectors are first superimposed and then passed to the network as input. The architecture and training parameters are optimized in an extensive automated hyperparameter optimization. The obtained segmentation results are statistically evaluated. Visualizations of the prediction confidence are used to identify existing problems. A key problem identified is the inadequate link between machine learning optimization targets and the microstructure parameters. The poster presents the latest results and shows the workflow for a grain boundary detection of a single-phase microstructure.