Page - 123 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

last step, the prepared steel specimen are scanned by a 4k line scan camera. The problem with conventional assessment approaches is that the preparation of the data material is very costly and time consuming. Thus, the available data material for this work consisted of only three blocks with manually annotated ground truth. III. RELATED WORK Vision-based approaches are already well established in assessment of material surface characteristics. As well there are also several approaches related to steel quality assess- ment. A computer vision based microstructure analysis and classification approach is introduced in [3]. The strategy is to set up a complex histogram representing a ’fingerprint’ of a microstructure. With the aid of those histograms it is possible to classify similar texture patterns by calculating the χ2 distance. Characterization of steel specimen surfaces are also pre- sented in [2].Signaturesof surfaceprofilesareextractedwith multiresolution wavelet decomposition. Furthermore, surface roughness parameters are derived from those signatures. Another feature extraction from micrographs is elaborated in [7]. The focus within this paper lies on extracting features like grain size, anisotropy of grains and the amount of δ phase. Further research on vision-based steel surface inspection mainly focuses on the detection of defects. A summary of detectable surface defects and approaches to identify them can be found in [5]. Nevertheless, the proposed methods focus on the analysis of microscopic scale specimens (few mm2) with their specific microscopicstructuresor thedetectionofdefects. Incontrast, the approach presented in this paper aims at the inspection and analysis of a full steel block with its macroscopic features. Those features exhibit completely different appear- ances than the microscopic structures. IV. QUALITY ASSESSMENT OF STEEL INGOTS Significant parameters for the quality of steel can be derived from so-called pool profiles, which can be derived from inspection of the remelted steel blocks. With the aid of those pool profiles it is possible to determine certain quality attributeswithin thewhole steelblock.Therefore theequality of the individual pool profile lines with their surroundings are taken intoaccount.Figure3showsmanuallyderivedpool profiles of an example steel block plate. These are generated by human experts (metallurgists) who try to identify the growth direction of the dendrites1 in the image. Based on those direction vectors, lines in predefined distances are estimated perpendicular to the vectors. This process is very time consuming and prone to human error. Furthermore, the results are influenced by subjective interpretation and, thus, experts easily end up with diverse results. 1Dendrites are complex three-dimensional tree-like structures. Dendritic morphology is the most commonly observed solidification structure [9], p. 78. Fig. 3: Manually derived pool profiles. Further ground truth data analysis revealed that some blocks show much more irregularities on top, bottom and in the middle due to the globular solidification in those areas. To be still able to extract meaningful pool profiles, metallurgists disregard those areas and simply classify pool profiles in regions with trans-crystalline solidification only. Thisbasicallymeans that trans-crystalline solidificationareas provide representative information, whereas globular areas are basically unstructured and as a result do not provide meaningful information for thepoolprofiles.Thus, foranob- jective evaluation it is essential to automatically distinguish between globular and trans-crystalline solidification areas. V. STEEL SPECIMEN SEGMENTATION The consequential first step of the automated quality as- sessment is the segmentationofglobular and trans-crystalline solidification areas. The main idea for automated segmenta- tion is based on the different textural appearance (regular and irregular patterns) of the different solidification regions. Therefore, various algorithms for the description of the surfaces were selected. The resulting classification gives information about where the actual extraction of information used for pool profile generation/calculation can be retrieved from. Due to the lack of extensive ground truth data, it was necessary to find suitable texture features and to implement customized classification methods rather than to train already existing classifiers. The following sections give an overview about the selected algorithms and the respective evaluation results. A. Gabor Filter The basic idea of using Gabor filters was to analyze spatial frequencies and their orientations within image patches. Trans-crystalline solidification areas represent areas with clearly visible frequencies and orientations whereas globular solidification areas do not. 2D Gabor filters are sinusoid functions combined with a Gaussian (see Figure 4) [6]. Two classes of training patches were created for globu- lar and trans-crystalline solidification areas. These patches 123