The measurement of grain size and grain size distribution is one of the most important quantification tasks in material science across a wide range of different materials, as it encodes information about materials properties as well as processing history. To automate this task, a large variety of machine learning approaches have already been developed for different and often highly complex microstructures, as modern Convolutional Neural Networks are potent enough to tackle the image analysis portion of the measurement. In practical use, the models are regularly confronted with a variety of hardships like imaging and preparation artefacts, changes in imaging conditions, substructures inside the grains, lack of reproducible etching conditions or even just weak contrast between grain and grain boundary phases. In other domains that require segmentation of fine objects or object boundaries, machine learning for contour completion has already been applied to improve segmentations results.
This work presents a two-step workflow to segment the grain boundaries in various challenging materials. In the first step, a non-material-specific semantic segmentation model is applied to the micrographs to obtain an initial grain boundary map. The second step applies a grain boundary completion model to iteratively identify and close incomplete grain boundaries that the first model failed to segment properly. Results are presented for a wide range of materials like FeNdB magnets, aluminum with Barker etching, prior austenite grains, copper, and others.