Bifurcated Attention and Feature Interaction for Few-Shot Fine-Grained Learning

Fine-grained image classification requires distinguishing visually similar categories with subtle semantic differences, yet traditional methods struggle with limited labeled data and computational inefficiency. To address these challenges, this study proposes a novel few-shot fine-grained classification framework integrating bifurcated attention and feature interaction mechanisms. The bifurcated attention module dynamically prioritizes critical regions through parallel spatial and channel attention pathways, enhancing discriminative feature extraction while reducing computational redundancy. A randomized query subset strategy optimizes parameter efficiency, and a feature interaction module adaptively aligns support and query features using cosine similarity and relation networks, improving cross-sample relationship modeling. Evaluations on CUB-200-2011, Stanford Dogs, and Stanford Cars datasets demonstrate state-of-the-art performance: achieving 5.95\% and 1.21\% accuracy gains in 5-way 1-shot/5-shot tasks on CUB-200-2011, and 4.15\% and 2.29\% improvements on Stanford Dogs. Complexity analysis confirms balanced memory usage and training efficiency, while visualizations validate the module's ability to capture long-range dependencies. This approach advances fine-grained classification under data scarcity by synergizing attention-guided feature refinement and metric learning, offering scalable solutions for real-world applications like rare species recognition and medical imaging.