Digital image processing and machine learning for automated microstructural characterization and grain size measurement in ferritic-pearlitic steels and nodular cast irons

Microstructural characterization of ferrous alloys, including ferritic-pearlitic steels and nodular cast irons, is fundamental for understanding and optimizing their mechanical performance. However, traditional manual or semi-automatic methods are often time-consuming, subjective, and lack reproducibility. This study aims to develop and validate a fully automated, robust methodology for simultaneous segmentation, classification, and grain/nodule size measurement of ferrite, pearlite, and graphite phases in multiphase ferrous alloys. A comprehensive pipeline was established, integrating FFT-based filtering and min–max normalization for image preprocessing, followed by Otsu thresholding and dual-mask watershed segmentation for grain and nodule isolation. After segmentation, morphological, intensity, and geometric features were extracted from each region. These features were then used to train and evaluate several supervised classifiers, including Random Forest, Logistic Regression, Support Vector Machine, and Neural Network. The models were validated using separate training and validation datasets, with images acquired under different imaging and etching conditions to ensure robustness and generalizability. The optimal model (Random Forest) achieved a classification accuracy of 96.96% across ferrite, pearlite, and graphite phases. Validation using our own metallographic images demonstrated accurate segmentation and classification of all phases, while the measured grain sizes and graphite nodule sizes were in close agreement with values reported by other authors in the literature. These results confirm the reliability of the method under different sample conditions. The proposed methodology provides a rapid, objective, and reproducible solution for quantitative microstructural analysis in both ferritic-pearlitic steels and nodular cast irons. By minimizing subjectivity and processing time, it offers substantial advantages for industrial quality control and advanced materials research, directly supporting structure–property–processing investigations.