<p>Fine-grained visual classification (FGVC) poses significant challenges due to subtle differences between classes and considerable variability within classes. While traditional convolutional neural networks (CNNs) excel at capturing local details, they often fall short in effectively modeling global information. In contrast, Transformer-based methods handle long-range dependencies but are computationally expensive and often struggle to focus on critical local regions. Research indicates that robust discriminative feature representation requires integrating both local and global features, making the primary challenge of FGVC the accurate differentiation of visually similar categories. To address this, we propose a Dual-Branch Multi-Feature Fusion Network (DBMFNet), which utilizes a Dual-Branch Feature Extraction Module (DFEM) to extract global features via Transformers while simultaneously capturing local details through CNNs. This architecture ensures the preservation of both modeling capabilities while maintaining spatial and channel consistency between global and local features, thereby providing structurally aligned feature representations that encompass global semantics and local details for subsequent fusion. To further enhance the representational capacity of these features, we introduce a Multi-Feature Fusion Module (MFM), which integrates channel attention, convolutional block attention, and an adaptive residual multilayer perceptron. This module effectively improves feature representations and facilitates adaptive fusion, resulting in the generation of more discriminative features. Experimental results demonstrate that DBMFNet achieves new state-of-the-art performance on the CUB-200-2011 (92.7%, +0.9%) and NABirds (92.4%, +1.0%) benchmarks, while delivering competitive performance on the Stanford Cars dataset (95.0%), showcasing its strong generalization and discriminative capabilities across diverse fine-grained tasks.</p>