Deep learning-powered system for automated detection and quantification of Vickers indentations

Hardness testing is a key procedure in materials science for evaluating mechanical properties and process quality. Traditional Vickers hardness measurement relies on manual identification of indentation diagonals, a process that is slow, subjective, and prone to variability. This work introduces a deep learning-based pipeline for fully automated Vickers hardness measurement, combining instance segmentation via Mask R-CNN with sub-pixel geometric fitting for diagonal extraction. A dataset of 403 micrographs of samples under loads from 10 gf to 2000 gf was assembled and annotated for training and validation. Hyperparameter optimisation was performed using a Taguchi design of experiments, and the final model achieved near-perfect segmentation accuracy (overall AP ≈ 90.5%) on the test set. Measurement accuracy was assessed against manual ground truth, yielding mean relative errors of 1.6-1.9% for the two diagonals, with most cases within 2-3%. These results demonstrate that the proposed system provides robust detection, high metrological precision, and reproducible performance across diverse imaging conditions, paving the way for reliable, high-throughput hardness testing in industrial and research settings.