Classification-guided two-stage segmentation of multi-feature M-A islands in bainitic steels

Abstract

The precise segmentation of multi-feature martensite–austenite (M–A) islands in bainitic steels is essential for understanding the microstructure–property relationships. However, current methods face challenges due to the complex morphology and interdependencies of features. This study proposes a classification-guided two-stage framework to address these issues, where the first stage employs classification to identify and categorize M–A islands, and the second stage performs segmentation using customized CBAM-enhanced U-Net models to accurately delineate their boundaries and capture detailed morphological features. Initially, a classification model categorizes M–A islands into distinct morphological types—elongated, blocky, and irregularly aggregated, providing essential prior knowledge for subsequent segmentation. Based on these classifications, customized segmentation models are developed for each type, optimized to enhance boundary accuracy and handle class imbalances. The results demonstrate superior segmentation performance across the different M–A island types, with average IoU values of 65.5 %, 88.0 %, and 81.5 % for elongated, blocky, and irregularly aggregated M–A islands, respectively. The proposed framework outperforms hybrid-model approaches while reducing reliance on large labeled datasets. In addition, a systematic evaluation of data increments demonstrates that the classification-guided strategy can achieve high accuracy with fewer annotations. Cross-material validation further confirms the strong generalization capability of the framework, underscoring its potential for broader applications in microstructural analysis. Overall, this study establishes a morphology-aware approach that enables precise and efficient microstructure classification and segmentation, while elucidating critical structure-property relationships governing fatigue-fracture resistance optimization in advanced steel systems.