Deep Transfer Learning for Automated Morphology Segmentation in vermicular cast iron

Automated microstructural analysis is important in materials science for objective characterization of phase composition and morphology. Recent deep transfer learning enables automated detection and segmentation in metallography. This study evaluates two deep learning approaches for detection, segmentation, and quantitative phase analysis in metallographic images. The first, based on YOLOv11, performs simultaneous detection and classification of ferrite, pearlite, graphite, and carbides, achieving good classification performance and consistent phase identification, but systematically underestimates phase fractions due to incomplete boundary segmentation. The second approach combines RT-DETR for detection with the Segment Anything Model (SAM) for precise segmentation, and a YOLO-based classifier for phase labeling. It provides accurate geometric segmentation and near-complete surface coverage, but classification performance is insufficient, reducing reliability. Comparative results show YOLOv11 favors classification over segmentation accuracy, while RT-DETR+SAM improves segmentation but weakens classification. These findings indicate the need for integrated methods combining classification and precise segmentation for automated microstructural characterization.