【摘要】 现有的基于图神经网络小样本分类方法，很少关注到与标签相关的语义信息。因此提出了基于混合语义的图神经网络小样本图像分类方法。使用补充词汇强化标签语义特征的表达能力，并将图像特征对齐到语义空间后，与标签语义特征混合生成实例级的混合语义特征。通过组合考虑任务上下文关系的图像特征和混合语义特征更好地描述样本，进而改善模型分类结果。在Mini-ImageNet和Tiered-ImageNet数据集上的实验结果表明，该方法对图像分类精度有明显的提高。更多还原

【Abstract】 Graph-based few-shot learning classification methods rarely pay attention to the semantic knowledge related to labels. This paper proposes hybrid-semantic based graph neural network for few-shot learning. Supplementary vocabulary is used to enhance the expression ability of label semantic features, and image features are aligned to semantic space and category semantic features are mixed to generate instance-level hybrid semantic features. By combining image features and mixed semantic features considering task context relations, the sample can be better described, and then the model classification results can be improved. The experimental results on Mini-ImageNet datasets and Tiered-ImageNet datasets show that the algorithm can significantly improve the image classification accuracy.更多还原