Grain size measurement based on a multi-supervised segmentation network and orientation fit closure method

Abstract The grain size rating is a key indicator for evaluating the microstructural properties of metallic materials. However, challenges such as the presence of twins, precipitates, and blurred grain boundaries complicate accurate and efficient grain size measurements. To address these issues, we propose a grain boundary segmentation network that combines a multisupervised signal network with multiple attention mechanisms. On the one hand, the integration of edge loss in the encoder makes the network focus more on the grain boundary regions. On the other hand, the multiple attention mechanism better integrates the low-level and high-level features of the image, enabling accurate grain boundary localization. Additionally, we design an orientation fit closure method to obtain closed-grain images. This method combines fracture boundary pixel path search with principal component analysis to determine the boundary extension direction. Finally, a quantitative analysis module is designed on the basis of GB/T 6394. The experimental results indicate that the rating results of the grain boundary images processed by the segmentation network and postprocessing are in close agreement with those provided by professional inspectors. The error for all test samples is within ±0.5 grades, providing objective and accurate data for the analysis of the physical properties of metallic materials.