EBSD image segmentation using a physics-based forward model

We propose a segmentation and anomaly detection method for electron backscatter diffraction (EBSD) images. In contrast to conventional methods that require Euler angles to be extracted from diffraction patterns, the proposed method operates on the patterns directly. We use a forward model implemented as a dictionary of diffraction patterns generated by a detailed physics-based simulation of EBSD. The combination of full diffraction patterns and a dictionary allows anomalies to be detected at the same time as grains are segmented, and also increases robustness to noise and instrument blur. The proposed method is demonstrated on a sample of the Ni-base alloy IN100.