Leveraging EBSD data by deep learning for bainite, ferrite and martensite segmentation

A U-Net model was trained to perform the segmentation of bainite, ferrite and martensite on EBSD maps using the kernel average misorientation and the pattern quality index as input. The manual labeling work was eased by introducing an "unknown" class that is ignored by the model during training. The influence of providing maps with different acquisition steps, indexation quality and constituent content to the model during training was investigated to demonstrate the importance of training the model with a wide range of configurations. The model can differentiate the three constituents with an 92% mean accuracy. An additional channel containing the map acquisition step was provided to the model and helped it generalize to various EBSD acquisition steps.