A mamba-architecture and multi-scale fusion enhanced artificial intelligent framework for precise microstructure segmentation from SEM image to EBSD labeling in medium Mn steel

Abstract

Medium-Mn steel (MMS) is crucial for lightweight automotive bodies. Accurately segmenting its retained austenite phase is key for property studies. While Electron backscatter diffraction (EBSD) possesses the ability to effectively characterize material information, it is expensive and time-consuming. Current deep learning techniques face challenges due to microstructural complexities, such as inhomogeneity and low contrast. To enhance the high accuracy of segmenting and quantifying the two-phase ratios, a Mamba-based multi-scale feature fusion semantic segmentation network (MFF-UMamba) is proposed, which integrates a Mamba framework with multi-scale feature fusion and deep supervision to enhance feature extraction. The experimental results of the proposed method on the medium Mn steel dataset achieved 79.02% in mean Intersection over Union (mIoU), reaching 88.41% in pixel accuracy, and 81.56% in Dice, with 5.47% mIoU improvement compared with the U-Net base model. The proposed method was applied to dual-phase steel dataset to verify the generality of MFF-UMamba, and mIoU was improved by 1.52% compared with the U-Net base model. Furthermore, the robustness of the model in different annealing parameters and the visualization effect of the network processing was also analyzed to explain the gain of the network on the dataset.