Segmentation of Three-Dimensional EBSD Data through Fast Multiscale Clustering

"Complete and accurate analysis of subgrain microstructural features must include three-dimensional information. Three-dimensional electron backscatter diffraction (EBSD) data can be used to characterize these features; however their boundaries first must be determined. This cannot be accomplished simply with pixel-to-pixel misorientation thresholding because many of the boundaries are gradual transitions in crystallographic orientation. Fast Multiscale Clustering (FMC) is an established image processing technique that we have combined with quaternion representation of orientation to segment this kind of data. Segmentation algorithms often have issues handling images with both distinct and subtle boundaries. Our implementation of FMC addresses this by using a novel distance function and statistical analysis to take into account the variance in orientation of each feature. Although FMC was originally a twodimensional image processing algorithm it can be generalized to analyze threedimensional data sets. As an example, a segmentation of microbands in cold-rolled aluminum is presented.IntroductionThree-dimensional electron backscatter diffraction (3D-EBSD) is a method for mapping the local crystallographic orientations of points in a volume. The technique uses a dual beam platform consisting of a focused ion beam (FIB) and SEM with EBSD, where FIB is used for serial sectioning and EBSD for orientation mapping of 2D slices. These 2D slices may then be stacked to create a 3D volumetric data set.Typical uses of such a data set include characterizing the geometries and crystallographic orientations of surfaces between structures in 3D [1-5]. This involves identifying boundaries accurately, either in 2D slices with subsequent 3D reconstruction (as illustrated in Figure 1) or in the 3D data set. All such data sets present the challenges of i) aligning data slices to correct for instrument drift, ii) identifying relevant boundaries, and iii) implementing appropriate smoothing or interpolation algorithms. The focus of this paper is boundary identification."