Indexing of electron back-scatter diffraction patterns using a convolutional neural network

Abstract

Accurate indexing of EBSD patterns presents a challenging problem. We propose a new convolutional neural network (EBSD-CNN) to realize real-time indexing of EBSD patterns; we implement a disorientation loss function to adapt a standard CNN model for crystallographic orientation indexing. The indexing accuracy, rate, and robustness against noise are evaluated using both simulated and experimental data, and compared with other indexing methods (Hough-based indexing, dictionary indexing, and spherical indexing). The results suggest that a CNN can provide an alternative to commercial Hough-transform-based indexing with comparative accuracy and rate. We obtain insight into the network functionality by visualization of selected filters.