||Almost 95% of the world steel production tonnage is solidified by continuous casting. The method is efficient, but is susceptible to crack formation. Cracks may cause problems later in the production chain, i.e. in the rolling mill, and should therefore be detected early in the process. An automated online crack detection system would be highly beneficial , but in order to ensure confidence in an automated inspection system false detection of cracks in non-defect regions must be avoided. An issue for a robust crack detection system is the presence of scales, which constitute a brittle oxidized top layer originating from the casting process. Cracks in the scales are frequently occurring, but must not be falsely detected as steel cracks.
In our work, 3D profile data of the steel surface, captured by laser triangulation, is used to scan casted steel slabs for longitudinal cracks. Segmentation of the data is done using mathematical morphology, and the resulting connected regions are assigned a crack probability estimate based on a logistic regression model.
Our current focus is on improvement of crack segment linking, i.e. how gaps in the crack signature can be bridged in order to increase the length of potential crack segments. Standard morphology operations rely on a predefined structuring element which is repeatedly used for each pixel in the image. The outline of a crack, however, can range from a straight line to a zig-zag pattern. A more adaptive method for linking regions with a large enough estimated crack depth would therefore be beneficial. More advanced morphological approaches, such as morphological amoebas and path openings, adapt better to curvature in the image. For our purpose, however, we investigate how the local structure tensor can be used to adaptively assign elliptical structuring elements based on the local orientation within the image. Similar techniques have previously been applied to assign orientation for line structuring elements, but require interpolation of directional fields in areas further away from strong directional features in the data. Instead, we let the information from the local structure tensor directly define the shape of the elliptical structuring element.