Fig. 5

An HfAl alloy image with dislocation defects is shown in (a) and model B predictions on the image are shown in (b). Model B was not trained on HfAl images. Green masks represent predicted lines, blue masks represent predicted loops, and teal represents overlapping predictions. This image has complex overlapping line defects as well as substantial noise that makes human annotation very difficult while model B had no difficulty making visually compelling defect quantifications. To quantify this, a subsection from the full HfAl image in (a) is further analyzed in (c) through (e). The highlighted subsection of the full image is shown in (c), the ground truth annotations of that subsection are shown in (d), and the cropping of the predictions in (b) for the subsection are in (e). The F1 score for lines in this subsection is 0.59 which is consistent with results found for other alloys analyzed in this study – see Table 2.