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 kernel average misorientation pattern quality index as input. The manual labeling work eased by introducing an “unknown” class that is ignored during training. influence providing with different acquisition steps, indexation constituent content training investigated demonstrate importance a wide range configurations. can differentiate three constituents 92% mean accuracy. An additional channel containing map step provided helped it generalize various steps.