计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (16): 111-122.DOI: 10.3778/j.issn.1002-8331.2012-0470

• 模式识别与人工智能 • 上一篇    下一篇

结合多尺度特征与掩码图网络的小样本学习

董博文,汪荣贵,杨娟,薛丽霞   

  1. 合肥工业大学 计算机与信息学院,合肥 230601
  • 出版日期:2022-08-15 发布日期:2022-08-15

Multi-Scale Feature Enhanced by Mask Graph Neural Network for Few-Shot Learning

DONG Bowen, WANG Ronggui, YANG Juan, XUE Lixia   

  1. School of Computer and Information, Hefei University of Technology, Hefei 230601, China
  • Online:2022-08-15 Published:2022-08-15

摘要: 对样本所含信息的提取能力决定网络模型进行小样本分类的效果,为了进一步提高模型挖掘信息的能力,提出一种结合多尺度特征与掩码图网络的小样本学习方法。设计由1×1卷积、全局平均池化和跳跃连接组成的最小残差神经网络块,与卷积块拼接成特征提取器,以提取样本不同尺度的特征,并通过注意力机制将不同尺度特征融合;使用融合的多尺度特征构建包含结点与边特征的图神经网络,并在其中加入一个元学习器(meta-learner)用于生成边的掩码,通过筛选边特征来指导图结点聚类与更新,进一步强化样本特征;通过特征贡献度和互斥损失改进类在嵌入空间表达特征的求解过程,提升模型度量学习能力。在MiniImagenet数据集上,该方法1-shot准确率为61.4%,5-shot准确率为78.6%,分别超过传统度量学习方法12.0个百分点与10.4个百分点;在Cifar-100数据集上分别提升9.7个百分点和6.0个百分点。该方法有效提升了小样本学习场景下的模型分类准确率。

关键词: 小样本学习, 度量学习, 元学习, 多尺度特征, 图神经网络

Abstract: The ability of extracting sample information determines the effect of the model for few-shot learning(FSL). To further improve the models’ ability of mining information, a method combining multi-scale feature with mask graph neural network(GNN) is proposed. Firstly, design themini residual network block composed of 1×1 convolution, global average pooling(GAP) and skip connection. Then, stitch mini residual network blocks with the convolutional blocks into feature extractors to extract features of different scales, and use the attention mechanism to fuse multi-scale feature. Secondly, a GNN containing node and edge features is constructed using the fused multi-scale feature. And a meta-learner is added into the GNN to generate edge masks to guide the clustering and updating of nodes by filtering edge features, which further strengthens the sample features. Finally, the computation of class representation is improved through the feature contribution and the pull loss, which enhances the ability of model metric learning. On the MiniImagenet dataset, the 1-shot accuracy of the method is 61.4%, and the 5-shot accuracy is 78.6%, which are 12.0 percentage points and 10.4 percentage points higher than that of the traditional metric learning method, on the Cifar-100 dataset, increases of 9.7 percentage points and 6.0 percentage points are reached respectively. Therefore, the method improves significantly the accuracy of few-shot learning based on metric learning.

Key words: few-shot learning, metric learning, meta-learning, multi-scale feature, graph neural network