Carbides, consisting of carbon and metal atoms with a particular crystal structure, are vital in materials science regarding phase transformation, microstructure evolution, and mechanical properties. Identifying the carbides from the microstructure matrix and analyzing the characteristics of carbides from a quantitative perspective is beneficial to facilitating the materials design. The deep learning method has been proven effective in identifying and quantifying carbide features; however, it should be based on large-sized training datasets. In this study, firstly we provided three brand-new datasets, i.e., carbides embedded in different microstructures, that have been manually annotated by materials experts. The mask R-CNN (Region-based Convolutional Neural Network) instance segmentation to these datasets as a benchmark was applied. Then, we used domain generalization theory and implement Cellpose, a deep learning-based segmentation technique that can accurately segment cells on images from a wide range of cell types. In our experiments, Cellpose’s generalization has been demonstrated in the domain of the steel microstructures. And this method is geared toward tasks with small training data or even no related training datasets, which can accurately separate cell-shape objectives from various image formats without requiring model retraining or parameter tweaks. A generalization of Cellpose in the context of steel structures were presented. To contribute to future research, we have re-configured the original datasets using Cellpose, which has allowed us to create a new dataset with more precisely defined borders. The auto segmentation of the carbides in the steel microstructures enables more accurate description of the precipitation features, which is of great benefit for various novel steel development in the future.