【摘要】 针对小样本图像分类问题，从卷积操作的局部连接性和基于非局部操作的注意力机制出发，提出了稀疏注意力关系网络(SARN)模型。在非局部操作过程中，利用稀疏策略筛选参与响应计算的相关特征。通过稀疏注意力机制构建不同空间位置相关特征之间的依赖性，切断语义无关特征之间的联系。后续卷积操作对不同空间位置的语义相关特征进行度量，抑制了无关信息的干扰，提高了模型的整体度量能力。通过在Mini-ImageNet和Tiered-ImageNet数据集上进行的一系列实验发现，相较于其他小样本学习模型，SARN模型的性能获得了显著提升。更多还原

【Abstract】 To solve the problem of small sample image classification, a Sparse Attention Relationship Network(SARN) model was proposed based on the local connectivity of convolution operations and the attention mechanism on the basic of non-local operations. In the process of non-local operation, the sparse strategy is used to calculate the relevant features involved in the response calculation. The dependence between the relevant features of different spatial locations is established through the sparse attention mechanism, and the connection of semantical irrelevant features is cut off. The subsequent convolution operation suppresses the interference of irrelevant information when performing feature measurement on semantical relevant features of different spacial positions, and improves the overall measurement ability of the model. Through a series of experiments on the Mini-ImageNet and Tiered-ImageNet datasets, it is found that SARN achieves significant performance improvement compared with small sample learning model.更多还原