Enhancing Scanning Electron Microscopy Analysis of Low-Carbon Steel: A Generative Adversarial Network-Based Approach for Image Standardization

The classification of complex microstructures in low-carbon steels is sensitive to imaging conditions, often causing domain shifts that degrade the accuracy of deep learning classifiers and confuse visual identification by experts. In this study, we constructed two SEM image datasets of low-carbon steels with eight heat treatments using field emission (FE) and tungsten (W) SEM sources. The accuracy of classifiers trained on images from one source and tested on images from another source showed a significant drop, from over 90% to around 40%. This finding underscores the significant impact of domain shift on both automated and visual classification. To address this problem, we used cycle-consistent generative adversarial networks (cycleGAN) to translate images between domains. This approach restored classifier accuracy to above 90% and successfully reproduced the distinct visual characteristics of each SEM source, confirming the effectiveness of cycleGAN in standardizing imaging conditions for reliable microstructural analysis.