Adaptive microstructure classification and segmentation of advanced high-strength steels using a two-stage deep learning framework

This study proposes a two-step deep learning framework for the automatic analysis of the microstructure of steels, including HSLA, DP, TRIP, MART, Q&P, and tool-grade steels. Convolutional neural network models were trained first to classify the steel grade, followed by the processing of the microstructure with a weakly supervised encoder-decoder-based CNN model for the segmentation of the phases. For this purpose, a steel microstructure dataset, comprising more than 2,400 images, was curated from published sources along with a few in-house-generated micrographs. The article discussed the results of processing a few typical microstructures with the proposed two-stage pipeline that includes the classification of grades and the segmentation of phases. The pretrained (with ImageNet weights) EfficientNetB6 was observed as an optimal architecture with an accuracy of 84.71% ± 2.33% on the test data, averaged over 5-fold cross-validation for the prediction of steel grade types. Nevertheless, in the second fold, processing of microstructures with the phase segmentation model shows an F1-score of more than 0.90 for the phases such as ferrite, martensite, bainite, retained austenite, and cementite. Overall, the proposed framework appears to be an efficient model to develop digital platforms for high-throughput automatic microstructure systems.