A few-shot fine-grained image classification method leveraging global and local structures

Few-shot fine-grained image classification aims to recognize sub-categories of the same super-category given only a few labeled samples. To deal with the low inter-class variation and the high intra-class discordance, both the supervised guidance from the global view and the detail information hidden in the local structure are necessary. However, such global structure and local detail are usually applied separately by existing methods, as a result, the features are not discriminative enough. To address this issue, we propose a novel few-shot fine-grained image classification framework which enhances the Discriminative ability of Local structures utilizing class-aware Global structures (DLG). Firstly, the DLG model calculates the global structures utilizing prototype representations of each class, and then constructs class-aware attention maps for query images to enhance their discriminative local structures with the aid of global structures. Finally, a classification module based on local structures is performed to make predictions. Results of case studies demonstrate that the class-aware attention maps can focus on class discriminative regions. Extensive experiments on fine-grained datasets demonstrate that DLG outperforms the state-of-the-art methods. Taking Stanford Dogs as an example, the proposed DLG outperforms the baselines. More specifically, DLG obtains at least 13.4% and 17.9% average gain on accuracy for 1-shot and 5-shot classification problem respectively. Code can be found at https:// gitee.com/csy213/few-shot-dlg.

Automated summary

This paper proposes a novel few-shot fine-grained image classification framework, DLG, which enhances the discriminative ability of local structures. Experimental results show that DLG outperforms the state-of-the-art methods on fine-grained datasets.