Imbalanced Classification of Ultrahigh Carbon Steel Microstructures with Vision Transformers and Image Synthesis of Minority Class

In the previous work., the authors have highlighted the potential of vision transformers., new types of deep learning models., in automatic microstructural classification of ultrahigh carbon steel., compared to convolutional neural networks. However., the scarcity of data associated with some classes or microstructures led to the underperformance of the deep learning models in the identification of these microstructures. In this work., as a follow-on to the previous one., the authors have carried out image synthesis of otherwise scarcely represented microstructures to address this class imbalance problem. Results have demonstrated that this additional data augmentation technique can reasonably alleviate the class imbalance problem and improve the model performance.