文献类型：期刊文献

中文题名：基于二阶统计量的小样本学习算法研究

英文题名：Research on Few-shot Learning Algorithm Based on Second-order Statistics

作者：麻永田[1];齐晶[2];张秋实[1];罗大为[1];方建军[1]

第一作者：麻永田

机构：[1]北京联合大学城市轨道交通与物流学院,北京100101;[2]北京联合大学旅游学院,北京100101

第一机构：北京联合大学城市轨道交通与物流学院

年份：2021

卷号：35

期号：4

起止页码：73-78

中文期刊名：北京联合大学学报

外文期刊名：Journal of Beijing Union University

语种：中文

中文关键词：小样本学习;协方差矩阵;二阶统计量;低维仿射;SVD分解

外文关键词：Few-shot learning;Covariance matrix;Second-order statistics;Low-dimensional affine;Singular value decomposition

摘要：为了提高小样本学习的准确率和抗干扰能力,提出了一种基于二阶统计量的小样本学习模型,以CNN最后一层卷积输出的一阶特征向量为输入,通过计算协方差矩阵和二阶池化获得具有较高区分度的二阶统计量,采用奇异值(SVD)分解将二阶特征映射到低维仿射子空间并据此分类。本算法在Omniglot和minilmageNet数据集上进行了测试,实验结果表明,在minilmageNet上的5-way 5-shot模型准确率达到了73.6%,比Prototypical Networks高出5.4%,在Omniglot上的20-way 1-shot模型准确率则获得了2.4%的提升,本算法性能优于Prototypical Networks等算法。在异常值测试中,本算法也展现出比Matching Networks和Prototypical Networks算法更强的鲁棒性。
To improve the accuracy and anti-interference ability of few-shot learning,this paper proposes a few-shot learning model based on second-order statistics.In the model,CNN is used to extract features and its output of the last convolutional layer is obtained to compute high-resolution second-order features by means of covariance matrix and second-order pooling operation.Meanwhile,the obtained second-order features are mapped to low-dimensional affine subspace by operating singular value decomposition(SVD)for classification.The proposed model is tested on Omniglot and minilmageNet datasets.The results reveal that the performance of the proposed model is better than other models including Prototypical Networks.The accuracy of the 5-way 5-shot model on minilmageNet dataset reaches up to 73.6%,which is 5.4%higher than Prototypical Networks.The 20-way 1-shot model on Omniglot dataset gets 2.4%accuracy improvement.As for outlier test,the proposed model also shows stronger robustness than those of Matching Networks and Prototypical Networks.