An automatic assessment scheme for steel quality inspection

This paper presents an automatic system for steel quality assessment, by me asuring textural properties of carbide distributions. In current steel insp ection, specially etched and polished steel specimen surfaces are classifie d manually under a light microscope, by comparisons with a standard chart. This procedure is basically two-dimensional, reflecting the size of the car bide agglomerations and their directional distribution. To capture these te xtural properties in terms of image features, we first apply a rich set of image-processing operations, including mathematical morphology, multi-chann el Gabor filtering, and the computation of texture measures with automatic scale selection in linear scale-space. Then, a feature selector is applied to a 40-dimensional feature space, and a classification scheme is defined, which on a sample set of more than 400 images has classification performanc e values comparable to those of human metallographers. Finally, a fully aut omatic inspection system is designed, which actively selects the most salie nt carbide structure on the specimen surface for subsequent classification. The feasibility of the overall approach for future use in the production p rocess is demonstrated by a prototype system. It is also shown how the pres ented classification scheme allows for the definition of a new reference ch art in terms of quantitative measures.