Open Access

Praanav Mahadev Menon

• This thesis presents the development and training of a U-Net- a convolutional neural network (CNN) model for the segmentation and classification of 4 microstructural phases in C45 steel: retained austenite, martensite, pearlite, and ferrite. It also features the steps taken to prepare the metallographic samples. The U-Net was designed and developed for biomedical microscopy imaging but has a wide array of uses for semantic segmentation in other fields as well. Utilizing Nital-etched microscopy images, the network was trained to identify and differentiate between these microstructures, which is critical for determining the mechanical properties of the steel. The dataset is comprised of a set of 35 grayscale images, and its corresponding labels/masks annotated with 4 colours for the various phases. Moreover, this thesis is an attempt at using deep learning to segregate retained austenite from martensite through visual LOM methods, as opposed to EBSD and XRD which are the widely used methods for identifying and quantifying retained austenite in a sample. The trained network demonstrated reasonable performance in segmenting the microstructural components. This research highlights the potential of deep learning approaches in materials science, particularly for automating the analysis of metallographic images, contributing to more efficient and precise material characterization.