Deep Learning Approaches for Automated Grain Size Estimation in Metallographic Analysis

The accurate classification of grain size in metallographic samples has long been a critical yet labor-intensive task for metallurgists. Knowledge of the correct grain size is extremely important, as different grain sizes result in varying mechanical, thermal and electromagnetic properties of the material. Traditional methods typically involve manual counting procedures that are time-consuming and prone to human error. Despite the availability of computer vision algorithms for grain size estimation, these solutions are often designed for idealized conditions, featuring clear edges and minimal noise, and frequently are not completely automated, particularly in the image preprocessing stage. A further complication arises from the wide variability in grain structures. In metallic samples, for example, heat treatments such as annealing can produce irregular shapes, defects, and blurred boundaries that challenge standard image-processing techniques. As well an improper specimen preparation can introduce image defects such as scratches as well as can do a dirty microscope lens. Moreover, to ensure the validity of grain size measurements, specific regulatory standards must be followed, such as the standard practices defined in ASTM E112-96, methodologies that the available algorithms often do not follow exactly. This thesis addresses the problem of fully automated grain size estimation in metallic materials, specifically focusing on AISI 304 and AISI 409 steels. It proposes a novel methodology that starts directly from raw microscope images and uses neural networks (NNs) to overcome the limitations of traditional computer vision approaches. By integrating image analysis with deep learning, this work aims to provide a robust and regulation-compliant system for automated grain size classification.