Page - 126 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

is clearly evident that areas with little or even no information content were masked out. The resulting orientations are smaller in number, but more expressive. Fig. 9: Optimization of orientation detection. Left: Result without filtering. Right: Result with filtering. C. Pool Profile Results The pool profile itself comprises of trace lines derived from ridge orientations. Each trace line is calculated from a given individual starting point by calculating a normal on the underlying orientation to the consequential next one and so forth. The calculation begins either from the outer borders (left and right) to the middle or vice versa. Figure 10 shows an example of automatic generated pool profiles overlaid on automatically detected orientations. The two colors of the trace lines represent the different starting orientations. Fig. 10: Automatic generated pool profiles from the sample steel block displayed in Figure 3. Metallurgists verified the quality of this approach by comparing the manually derived ground truth (Figure 3) with the achieved results (Figure 10). The comparison shows the good correspondence of manually generated ground truth with automated derived pool profiles. VII. CONCLUSIONS AND OUTLOOK This paper presented algorithms to perform steel speci- men segmentation for classification of globular and trans- crystalline solidification areas and algorithms to automate pool profile generation. Figure 11 displays a whole steel block with segmentation and pool profiles. The automated quality assessment is currently under evaluation by metallur- gists on additional steel blocks. Fig. 11: Final result with segmentation and pool profiles. 126