Enhancing few-shot fine-grained image classification via adaptive attention and task-specific modulation

Abstract The few-shot fine-grained image classification task faces two key challenges: first, the difficulty in detecting subtle inter-class differences under limited labeled samples; second, significant intra-class variance leading to poor feature generalization ability. To address these issues, this paper proposes a novel Few-Shot Fine-grained Classification Network (FecNet), which achieves efficient feature learning through the Multidimensional Collaborative Attention (MCA) and Task Adaptive Modulation (TAM) modules. The MCA module innovatively integrates attentional inference in three dimensions (channel, width, and height), significantly improving the discrimination ability of features while keeping computational complexity at only 0.8%. The TAM module aligns local key features of the support set and query set through cross-sample 3D convolutional correlation extraction, effectively reducing inter-class distances. Experimental results on three benchmark datasets (CUB-200-2011, Stanford Cars, and Stanford Dogs) show that FecNet based on the Conv-4 backbone network achieves a classification accuracy of 71.23% on Dogs and 84.01% on Cars in the 5-way 1-shot task, outperforming existing methods by 3.9% and 2.72% respectively. Ablation experiments confirm the synergistic effectiveness of the MCA and TAM modules. This study provides a new solution for few-shot fine-grained classification and has important application value in practical scenarios such as biometrics.