Key points are not available for this paper at this time.
Artificial intelligence (AI) has rapidly become a key enabler in materials science, facilitating automated, high-precision analysis of microstructure and surface characteristics. This review presents a structured overview of both deep learning (DL) and classical image processing approaches used in tasks such as segmentation, classification, super-resolution, and surface defect detection across different imaging techniques, including optical microscopy, scanning electron microscopy (SEM), atomic force microscopy (AFM), and electron backscatter diffraction (EBSD). The reviewed studies show that modern DL architectures such as convolutional neural networks (CNN), U-Net, Vision Transformers, and YOLO achieve high performance, often reporting accuracy levels above 90% in segmentation and defect detection tasks. In addition, generative models improve robustness by enabling data augmentation and the generation of synthetic data, particularly in data-limited scenarios. Despite these advances, important challenges remain, including limited data availability, domain shifts across imaging conditions, and limited interpretability of deep models. This review therefore examines explainable AI (XAI) approaches and highlights that visual explanations alone are insufficient unless supported by physical microstructural knowledge. At the same time, classical feature-based methods such as GLCM, LBP, and Gabor filters continue to offer advantages in interpretability and computational efficiency, especially in low-data scenarios. Overall, the review emphasizes the complementary use of classical and deep learning approaches and outlines future directions, including standardized datasets, domain adaptation, and physics-informed modeling to improve reliability and generalization in AI-driven materials characterization.
Erçetin et al. (Sat,) studied this question.