The task of Fine-Grained Visual Classification (FGVC) aims to distinguish between closely related subclasses within a broader category. Challenges include high intra-class variation, minimal inter-class differences, and complex backgrounds, making accurate classification demanding. Existing approaches usually target on specific traits for optimization and do not pay enough attention to both local details and the object’s structural information in images simultaneously. In this paper, We designed a new Object Region Attention transformer (ORA-Trans), which enables the model to better discern subtle image differences by mining deeply into discriminative regions and structural features of objects in images. Specifically, we propose a Key Tokens Selector to automatically identify tokens corresponding to crucial local regions that are pivotal for FGVC. Furthermore, we design an object Structure Feature Extractor to capture the structural relationships among the object parts tokens, which can not only optimize the feature learning process of the tokens selected by KTS, but also alleviate the interference of complex image backgrounds. We also introduce a Token Stochastic Swap module to enrich feature combinations and enhance feature representativeness. Our proposed approach achieves state-of-the-art performances on the CUB_200_2011, Stanford dogs, and IP102 datasets. Ablation studies and visualization demonstrate the effectiveness and interpretability of our method.