Workflow Consisting of DNN-Based Segmentation Method and Persistent Homology Analysis for Feature Extraction from Microstructural Images

It is becoming increasingly important to use data on materials in order to accelerate materials development, particularly data on the microstructure of materials because they are associated with both processes and characteristics. The data typically acquired for microstructures are microstructural images; however, it is inefficient to use them directly for materials development because the microstructural images are two-dimensional unstructured data. Therefore, it is necessary to apply image analysis, but this also presents some challenges.<br/>The first issue relates to the workload for image analysis. To analyzes the microstructural images, it is necessary to determine the region of interest (ROI) which is time-consuming and requires a highly skilled analyst. The second issue is the feature extraction from the microstructural images. Although various extraction methods have been proposed, extracting the features of the morphology of the ROI is still considerably challenging.<br/>In this study, we developed a workflow that consists of deep neural network (DNN) based segmentation method [1] and persistent homology analysis [2] to solve the above problems. The DNN-based segmentation method was utilized to solve the workload problem by automatically converting the microstructural images to segmented images. The persistent homology analysis was utilized for feature extraction by expressing the geometric shape and intermingling characteristics of ROI as a persistence diagram. The persistence diagram can extract the morphological characteristics of both connected and isolated regions.<br/>The workflow was applied to a Cr-based alloy with a complex structure. Sample of the alloy were prepared by cutting from a cylindrical ingot that was formed by pouring and cooling molten metal in a mold. The exterior of the ingot cooled rapidly while the center cooled gradually. The microstructure at three positions (center, edge, midpoint of the center and the edge) in the ingot were observed by using a scanning electron microscope and obtaining 50 field of view backscattered electron microscopic images. The images were automatically converted to segmented images by applying the DNN-based segmentation method. Then the persistent homology analysis was carried out to extract the features of the ROI as a distribution of plots in the persistence diagram. The persistence diagram from the 50 field of view images was integrated to reduce the cross-field variation. We defined new feature points in the persistence diagram distribution to investigate the relation between the cooling rate and the thickness of the ROI on the microstructures. The new feature points revealed that the difference in the microstructures of the sampling points were caused by differences in the cooling rate.<br/><br/>[1] V. Badrinarayanan, A. Kendall, and R. Cipolla, IEEE Trans. Pattern Anal. Mach. Intell. <b>39</b>, 2481-2495 (2017).<br/>[2] Edelsbrunner et al., Discrete Comput. Geom. <b>28</b>, 511–533, (2002).