Feb/March_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 | F E B R U A R Y / M A R C H 2 0 2 1 1 5 images with feature vectors closest to that of a given target image; obvious- ly, feature vector similarity is reflected in visual similarity. This makes it easy to search an image database for relat- ed microstructures. For the same set of UHCS micrographs, Fig. 2 shows a vi- sual clustering map where each point represents an image; point color corre- sponds to the primary microstructural constituent in each micrograph. Clearly, similar images cluster, which illustrates the visual structure of the data set. Feature vectors can also be used to quantify microstructural information directly. Figure 3 shows an example that uses the feature vector to measure av- erage grain size in polycrystalline mi- crostructures; the results are within a standard error of 2.3%. Finally, the fea- ture vector can contain visual informa- tion that is not perceptible to humans. For instance, chemical composition is not usually measured visually, but rath- er with specialized tools such as energy dispersive spectroscopy (EDS). Figure 4 shows the results of a ML approach that achieves 76% total accuracy in classi- fying the composition of inclusions in steel from backscattered SEM images. This demonstrates the ability of the CV/ ML system to sense subtle visual details like feature size, shape, contrast, and color distribution with a fidelity that ex- ceeds human perception. SEMANTIC SEGMENTATION Quantitative measurement of ma- terials microstructure typically requires the image to be segmented, where each pixel in the image is assigned to a mi- crostructural constituent. Convention- al segmentation algorithms, such as those incorporated in ImageJ [19] , can work well on suitable microstructures, but become less effective for com- plex or non-ideal images and often re- quire considerable human intervention. Therefore, we turn to CV/ML methods to address these challenges. Image segmentation has import- ant applications in robotics andmedical imaging among others, so there is con- siderable researchactivity indeveloping segmentation methods. These meth- ods can be adapted to microstructural Fig. 1 — Visual search for images with similar feature vectors in a database of 961 ultrahigh carbon steel micrographs [18] . Visually similar micrographs contain similar microstructural constituents, here comprised of a mix of network, Widmanstatten, and spheroidite carbides. Fig. 2 — Visual clustering plot in which micrographs cluster according to their primary microconstituent: spheroidite (blue), network carbide (red), or pearlite (green) [2] . Example microstructures for each cluster are shown in the insets.
Made with FlippingBook
RkJQdWJsaXNoZXIy MjA4MTAy