January_2021_AMP_Digital

A D V A N C E D M A T E R I A L S & P R O C E S S E S | J A N U A R Y 2 0 2 1 1 8 as well as the fineness and complexity of the structures. Also, there is no con- sensus among human experts about microstructure formation mecha- nisms [10] nor in labeling and classifying bainitic structures [4] . Thus, it is ideally suited for this case study. A simple and effective means for a more objective ground truth assign- ment are so-called round robin tests in which images are given to a group of ex- perts. Every expert judges the image in- dividually. In the end, the image’s class will be determined by a majority vote, i.e., the class assigned by most experts will be ultimately chosen. However, for bainitic structures this method can reach its limit when there is too much disagreement between experts and no preferred class. Then it is the task of materials scientists to find other meth- ods that do not purely rely on how the microstructure visually appears to the expert eye. The following sections will look at three approaches: (1) correlative microscopy using electron backscat- ter diffraction (EBSD) as an additional information source; (2) use of specifi- cally produced reference samples; and (3) unsupervised learning techniques. CORRELATIVE MICROSCOPY Usually, one single characteriza- tion method cannot capture all micro- structural features relevant for quanti- fication. By combining different char- acterization methods in a correlative approach, features fromdifferent length scales and varying complementing in- formation sources can be combined, thus overcoming the disadvantages of a particular method. In this correlative approach, standard light optical mi- croscopy (LOM), scanning electron mi- croscopy (SEM), and additional EBSD measurements are combined. While LOM and SEM “only” enable the visual inspection of the microstructure, EBSD provides a completely different set of information, e.g., misorientation pa- rameters or data about both grain and phase boundaries, which—in the case of bainite in particular—ideally com- plement the LOM and SEM images and help to better understand and assess the microstructure. This is exactly the added value that EBSD and correlative approaches in general provide. Further- more, there is always the chance that the knowledge and references gained with the additional methods in the cor- relative approach (here EBSD), may per- mit to reduce further investigations to only one characterization method, ide- ally the simplest (here LOM). To fully exploit the data from these different, complementing information sources, the different micrographs must be registered. For the methodology of image registration, the authors refer to the references [10,11] . Figure 3 shows two examples of how combining EBSD pa- rameters with LOM and/or SEM images by overlaying can help in assigning the ground truth. Just relying on the LOM and/or SEM image may result in varying results from different experts. In the case of the complex-phase steel micrograph in Fig. 3a, the lath- like bainite regions are to be annotated for ML segmentation. EBSD parameters Fig. 2 — Comparison of typical characteristics of natural scene (a) and microstructural images (b).

RkJQdWJsaXNoZXIy MjA4MTAy