Seite - 125 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

Fig. 7: LBP dominating feature/bin output. Figure 7 shows a color coded image on the left hand side where each color matches a bin from the summarized histogram. The output image on the right hand side was generated by using a majority filter calculating the dom- inating feature for a specified area around a center pixel and then plotting its assigned color in the output image. The background colors of the illustrated patterns (in the middle) correlate with the colors in the left image. Together they represent the orange area within the final binary output image on the right. It is clearly evident that horizontal lines (line endings) smoothly correlate with globular solidification areas whereas trans-crystalline areas are dominated by other orientations. D. Feature Comparison Experiments with different steel compositions have shown that the Gabor, as well as the spatial filter bank approach, do not deliver generic solutions. Even for equal types of steel with other block dimensions, those algorithms do not deliver satisfying results. Interestingly, the discovery that dominating horizontal orientations correlate with the globular solidification area, wasalso proven for further steel blocks.The validation of the segmentationoutputwasperformedbymetallurgistsvisually. Thus, the segmentation based on LBP (Figure 7) is used as basis for the quality parameter extraction. VI. POOL PROFILES As previously mentioned, pool profiles are used to de- termine quality parameters. Therefore, a fast and reliable process that can produce repeatable results with a minimum need of human interaction is required. The best performing method during analysis of different steel types is based on a combination of scale-space and ridge detection. The ridge detection is similar to a biomet- ric fingerprint recognition approach [4] with the difference that in this application regions with constant directions are important, whereas in fingerprint recognition characteristics like crossing points or ridge ends are relevant. A. Ridges and Orientations The algorithm for ridge preparation, extraction and orien- tation calculation is based on a paper presented by Hong, Wan and Jain [4]. They show a way to identify and nor- malize ridge regions within an image and to calculate their orientations. Figure 8 illustrates the process of pool profile derivation on a small sample sector of a steel specimen. Fig. 8: Pool profile detection by ridge analysis. Top: Sector of steel specimen and derived ridges. Bottom: Derivation of orientations and final pool profile. B. Orientation Filtering The cutting or etching process in preparation of the steel sample or the imaging/scanning process itself can lead to artifacts. In order to handle those problematic areas, it is necessary to implement a filtering algorithm for the ridge orientations. The first step of optimization takes place during pre- processing and delivers a mask of non-valid areas through a gray scale segmentation process performed on a smoothed and re-sampled image of the steel specimen. The second step is the filtering of derived orientations. This filtering relies on homogeneity properties of the ori- entations in image areas. If an orientation vector is not in conformity with its surrounding/neighboring orientations it is treated as outlier and, thus, filtered before deriving the final pool profile. Additionally, as a third step orientations are calculated on different scales of the image to pre-filter orientations deviating from smaller scales. Figure 9 shows a sample steel plate and the calculated orientations (short blue lines) with and without filtering. It 125