Where Windows Meet State Spaces: A Hybrid Network for Fine-Grained Image Recognition

Image classification is one of the main tasks of computer vision. However, the multi-class recognition has not gained much traction in many visual domains due to limited long-range modeling and high computational costs that often lead to poor generalization. This paper develops a cross-attentive Swin–Vision Mamba hybrid network with hierarchical window attention and state-space sequence modeling. By using dual-branch framework to capture local textures and global context and fuse them, Cross-attention and gated aggregation are used. We applied advanced augmentation, label smoothing, and cosine scheduling for training, and evaluated on Caltech-256 and other cross-domain datasets. Our experiments comparing results with modern baselines such as ViT, Swin V2, ConvNeXt V2, and Vision Mamba achieve accuracy, macro-F1, robustness, and calibration improvements while keeping computational cost competitive. Ablation studies confirm that attention and state-space branches complement each other. The framework established in this research is reproducible, efficient, and generalizable for various visual recognition tasks, which allows for strong accuracy–efficiency trade-offs and design principles for hybrid architectures.