EXTENDED ABSTRACT: The rapid growth of machine learning has led to its widespread application in materials science, enabling deep analysis of material images. This approach uncovers the intrinsic relationships between microstructure and material properties, playing a crucial role in predicting performance and developing new materials. To address challenges posed by small sample sizes and complex microstructures in image segmentation, this study proposes a dual-branch semantic segmentation network based on feature pyramids and cross-attention. The primary branch aggregates multi-level features to enhance detail, while the auxiliary branch focuses on learning texture and boundaries. Statistical analysis of different individuals within the same category is conducted based on segmentation results. For instance segmentation, the study introduces a multimodal fusion and pseudo-labeling approach to improve data utilization. Using duplex stainless steel, the distribution of the secondary σ phase was determined, with its coarsening rate calculated at 35.3491 nm/s^(1/3) and the interface energy between the σ phase and γ phase estimated at 1.708x10^4 J/m^2 according to the Ostwald ripening mechanism. To improve feature extraction and address small sample limitations, a dual-branch multi-scale strategy was developed, proposing a global-local feature focus network. This network independently extracts global and local features, which are then fused through a global-local feature fusion module. A high-dimensional feature abstraction and clustering module refines these features, reducing the model’s dependence on large datasets. The method processes various modal inputs, such as element distribution and microstructure. An element-enhanced strategy helps the model better understand microstructures, leading to more accurate performance predictions. This study applied the method to YSZ coatings doped with Al2O3, achieving high prediction accuracy for thermal and electrical conductivity, with R2 values of 0.974 and 0.932, respectively. For hypoeutectic cast Al-Si alloys, a semantic segmentation model combined with a contour-based parameter extraction method enabled pixel-level segmentation of eutectic Si images and automatic extraction of microstructural parameters. This data was used to construct a regression model to predict tensile strength based on composition and microstructure. The methods proposed in this study are versatile and can be applied to the prediction and analysis of various material properties, offering a robust approach to handling material images.

Keywords: Small sample; material image processing; material image-performance mining; machine learning