Classi cation of Carbide Distributions using Scale-Space Methods Classi cation of Carbide Distributions using Scale Selection and Directional Distributions

In the production of high speed steel, the rolling a ects the micro-structure of the steel, which in turn in uences the mechanical properties. Speci cally, the distribution of carbide is essential, since cracks propagate within the carbide agglomerations. In current quality control, the properties of the steel are assessed manually by comparison with a standard chart, containing representative patterns for each steel class. Interestingly, the standard technique for classifying carbide distributions is two-dimensional, where the rst dimension basically corresponds to scale (\degree" | the size of the largest carbide agglomeration) and the the second dimension basically re ects the directional distribution (\type" | how strongly the net structure of carbide has been stretched). In this paper, we present an automatic method for such classi cation based on scale-space operations, in which the size information is measured using recently developed techniques for feature detection with automatic scale selection and the directional information is computed from secondmoment descriptors (Lindeberg 1994). Combined with a morphological veri cation scheme, a pattern classi er is proposed, which shares large similarities with current manual techniques. Compared to previous work (Wiltschi, Pinz & Hackl 1995), the proposed scheme has the advantage that the signi cant scale of the carbide agglomeration is calculated explicitly, and the method is much less sensitive to the variance of spatial connectivity than a morphological approach. From a theoretical viewpoint, the proposed scheme also has the attractive property that it is based on similar visual-front-end operations as a large class of computer vision modules.