Indexing high-noise electron backscatter diffraction patterns using convolutional neural network and transfer learning

Abstract

The crystal orientation indexing method based on Hough transform has many “zero solution” points when indexing the crystal orientation of materials, especially those with large strains. We propose an efficient and migratable indexing method for fast and autonomous crystal orientations indexing of experimental electron backscatter diffraction (EBSD) patterns of various crystal structures by transfer learning. To better evaluate the performance of the model, the performance of the model was assessed both quantitatively based on the numerical metrics and qualitatively through the Inverse Pole Figure (IPF). The results of indexing high-noise nickel EBSD patterns show that the CNN-based indexing method could not only obtain an acceptable indexing accuracy (a mean disorientation error of 1.77° for retraining and 2.31° for transfer learning), but also reasonably index the “zero solution” points remained in the Hough transform-based indexing method. And the results of indexing other materials by transfer learning show that the CNN has strong knowledge migration capability. In addition, the interpretability of the CNN was investigated by visualizing feature maps of each convolutional layer, and the sensitivity analysis of the CNN was realized by testing masked images.