Prior austenite grain boundary recognition in martensite microstructure based on deep learning

Grain size determination is essential in producing and testing iron and steel materials. Grain size determination of martensitic steels usually requires etching with picric acid to reveal the prior austenite grain boundaries. However, picric acid is toxic and explosive and belongs to hazardous chemicals, which makes it difficult for laboratories and testing institutions to obtain. A new experimental method was developed to use Nital etchant instead of picric acid. The deep learning method was used to recognize the prior austenite grain boundaries in the etched martensite microstructure, and the grain size could be determined according to the recognition result. Firstly, the polished martensite specimen was etched twice with Nital etchant and picric acid, respectively, and the same position was observed using an optical microscope. The images of the martensitic structure and its prior austenite grain boundary label were obtained, and a data set was constructed. Secondly, based on this data set, a convolutional neural network model with a semantic segmentation function was trained, and the accuracy rate of the test set was 87.53%. Finally, according to the recognition results of the model, the grain size rating can be automatically determined or provide a reference for experimenters, and the difference between the automatic determination results and the measured results is about 0.5 level.