Computer vision has recently made notable progress, with deep learning-based approaches demonstrating outstanding performance in visual classification tasks. Fine-grained visual classification (FGVC) still faces challenges due to high inter-class similarity and substantial intra-class differences. Vision Transformer (ViT), leveraging the self-attention mechanism, has demonstrated impressive performance in FGVC. Nevertheless, further performance enhancement necessitates refined processing of local features. To tackle this issue, this paper introduces an improved ViT-based approach. A Dynamic Local Feature Enhancement Module is introduced into the ViT backbone, which adaptively focuses on discriminative regions using deformable attention and enhances local feature representations by incorporating frequency-domain channel attention. Second, Adaptive Attention Feature Fusion Module is designed, integrating CBAM to achieve complementary learning between global and local features. Finally, ArcFace Loss is incorporated alongside conventional cross-entropy loss to optimize class boundaries.