Determining crystallographic orientation via hybrid convolutional neural network

Highlights: • CNN accelerated hybrid orientaiton indexing (EBSDDI-CNN) is both accurate and robust against noise. • Performance of EBSDDI-CNN can be fine-tuned by adjusting the Top_K value. • EBSDDI-CNN reduces computational time of DI by two-thirds. • The acceleration of EBSDDI-CNN is orientation-dependent. A recent paradigm shift in the electron diffraction community has benefited from accessibility of large data sets and ever more complex designs of convolutional neural networks (CNNs). However, this shift from conventional feature engineering to analyzing high-level features extracted from CNN is often accompanied by a reduction in accuracy and sensitivity. Particularly, CNN based crystal orientation indexing using electron backscatter diffraction is sensitive to noise, reducing the overall accuracy. In this study, a new hybrid indexing approach has been developed to integrate dictionary indexing (DI) with a trained CNN to achieve extraordinary speed and robustness against noise simultaneously.