Non-metallic inclusions (NMI) are phases such as sulfides and oxides, which are distributed in the steel matrix and are characterized by different shapes and sizes. Quantitative microstructure analysis of such inclusions is important to study their effects on steel properties. Existing approaches classify inclusions based on the color and morphological characteristics but tend to misclassify sample preparation artefacts such as stains or scratches. If new artefacts appear, their characteristics have to be updated.
A new approach was developed, based on a supervised machine learning (ML) model, trained on texture features of the segmented objects. Using a random forest classifier, trained on Haralick parameters of inclusions and artefacts, it is possible to differentiate artefacts even if they have similar shape characteristics as those of the NMI.
The goal is to develop a machine learning model to differentiate NMI from artefacts without defining all possible artefacts in the sample a priori. The texture patterns of NMI help the model to identify new artefacts with greater accuracy because it is independent of morphological characteristics. Such models are more robust to micrographs of not perfect quality. In addition, inhomogeneous illumination (shading) can be compensated. With this machine learning based approach a more accurate, robust and time efficient analysis of non-metallic inclusions in steels is possible.