基于改进关系网络的小样本学习

Few-shot learning using improved relation network

关系网络是一种端到端的小样本学习框架,可以通过少量标注样本识别新的类别.在关系网络的基础上,融合inception块和感受野块,提出了一种基于改进关系网络的小样本学习方法.用inception块替换嵌入模型的第3个卷积层,并且将感受野块添加在关系模型的起始位置,这两种卷积块分别提升了网络的特征表达能力和度量能力.在Omniglot数据集上,该算法的识别率整体高于关系网络,达到97%以上;在miniImagenet数据集上,采用5-way 1-shot和5-way 5-shot方法,算法识别率分别达到52.89%,67.15%.

Relation network was an end-to-end few-shot learning framework,which classified new classes with few labeled samples.Based on relation network,combined inception block and receptive field block,few-shot learning method using improved relation network was proposed,which replaced the third convolution layer of the embedding module with the inception block,and added the receptive field block to the starting position of the relational module.These two types of convolutional blocks improved the network’s feature expression and metric ability.On Omniglot,the accuracy of few-shot learning method using improved relation network was higher than that of relation network,reached 97%.On miniImagenet,the 5-way 1-shot and 5-way 5-shot methods arrived 52.89%and 67.15%.