In microstructure analysis, the precise identification of grains and grain boundaries for determining grain size and shape is a tedious and time-consuming manual task that also requires expert knowledge. Existing automated techniques, such as edge detection methods, are considerably faster than manual analysis but often suffer from limited accuracy. Accurate prediction of grain-related properties is essential in several fields, including understanding material properties, predicting material behavior, quality control, and process optimization. To address these challenges, deep learning-based convolutional neural network (CNN) models can be effectively employed. In the present work, a novel method is proposed for detecting grain boundaries and extrapolating incomplete boundaries to form closed grain regions, thereby enabling accurate quantification of grain parameters. This study introduces an ensemble learning–based framework comprising two models: the first employs a unique progressive learning strategy, while the second adopts an iterative approach to close incomplete grain boundaries. The proposed framework is implemented using the U-net architecture. A dataset of microstructure images is acquired using an optical microscope at varied magnifications. The materials investigated are mild steel and Inconel 625, prepared through standard metallographic procedures including polishing and etching prior to image acquisition. The ensemble model is trained on a diverse dataset that also includes artificially generated images. The developed model demonstrates high accuracy in predicting several grain characteristics, including average grain area, total number of grains, grain height, and grain width. The results indicate that the network effectively learns these properties by capturing the curvature and morphology of grain boundaries.

Keywords

Grain boundary prediction, Grain size, CNN, U-Net architecture, Ensemble Learning, Quantitative analysis