FSCA: Few-Shot Learning via Embedding Adaptation with Corner Multi-Head Attention

The demand for computer vision applications has been growing in recent years. However, most tasks in real-world scenarios are in the state of small amounts of data, and the discriminability of features results in low performance in few-shot settings. In order to obtain discriminability features in unknown classes more robustly, we first propose a simple and effective few-shot feature augmentation scheme called corner multi-head attention (CMA). The CMA module aims to project corner features through multi-head attention into a task-specific feature space through corner detection without sacrificing the generalization ability of the few-shot model. We then construct a representation space for few-shot classification by adaptively embedding the skeleton (EA). Finally, we proposed a few-shot classification framework, namely FSCA (Few-Shot Embedding Adaptation with Corner Multi-Head Attention). Its CMA module has low coupling, supports plug-and-play, and can easily be embedded into various models. To verify the effectiveness of the proposed scheme, we conducted experiments on three few-shot classification data sets, CUB-200-2011, miniImageNet, and tieredImageNet, and we performed a t-SNE clustering visual analysis. The experimental results show that it achieves continuous improvements over state-of-the-art methods on three widely used few-shot classification benchmarks. Specifically, on the miniImageNet data, the performance of FSCA under the five-way, one-shot and five-shot settings improved by an average of 1.6–4.5% and 2.2–5.4%, respectively.