Segmentation and classification of clinker crystals in microscopic images by statistical learning methods

The quality analysis process of Portland cement is currently carried out by trained professionals who analyze the crystals present in the microstructure of the clinker (an input produced in the cement manufacturing process and which gives it its main characteristics). Among these crystals, the one that most affects the quality of the final product is the alite ( $$\varvec{C_3S}$$ ). Because of this, building an automatic process for segmenting and classifying $$\varvec{C_3S}$$ in microscopic images of clinker can bring savings and efficiency in cement manufacturing. This work, therefore, seeks, through convolutional neural networks, data weighting for balancing, image preprocessing filters and other statistical learning methods, to carry out this segmentation and classification so that the automation of the process is viable, enhancing the monitoring of the quality of the product. Several statistical learning methods combined with different preprocessing filters and data weighting are trained and compared in the analysis of clinker images. Preprocessing filters did not significantly improve the segmentation and classification process, but balancing classes via weights promoted a representative improvement. Classifying the crystals in a second step, after image segmentation, using simpler statistical learning methods than convolutional neural networks and utilizing crystal shape characteristics proved more efficient than classifying them along with segmentation.