摘要
当前深度学习大都基于大量数据通过构建深层次的网络实现自动识别,但在很多场景中难以获得大量的样本数据。针对这一问题,提出一种基于孪生变分自编码器(siamese variational auto-encoder,S-VAE)的小样本图像分类方法。通过变分自编码器提取原始训练数据的高层语义特征,然后由两个训练好的变分自编码器的编码器部分组建孪生网络的输入结构,最后通过分类器对样本进行识别。变分自编码器可以解决样本数据量少带来的过拟合问题,孪生网络的结构增加了样本数量较少的情况下的训练次数。在Omniglot数据集上进行的实验结果表明:本方法与原始孪生神经网络相比正确率平均提高了3.1%,模型收敛速度更快,证明了孪生变分自编码器能够较好地完成小样本数据分类任务。
Currently,most deep learning is based on the use of large amounts of data and the construction of a deep network to achieve automatic recognition,but it is difficult to obtain a large amount of sample data in many scenarios.To solve this problem,we propose the use of a small-sample image classification method based on a Siamese variational auto-encoder(S-VAE).First,the high-level semantic features of the original training data are extracted by the variational auto-encoder.Then,the input structure of the Siamese network is constructed by the encoders of two trained variational auto-encoders.Lastly,the samples are identified by the classifier.The variational auto-encoder can solve the over-fitting problem caused by small amounts of sample data,and the structure of the Siamese network increases the training number when the sample size is small.The experimental results on the Omniglot dataset show that this method has a 3.1%average improvement in accuracy compared with the original Siamese network.The model convergence speed is also faster,which proves that the S-VAE can better complete the classification task when using small-sample data.
作者
王德文
魏波涛
WANG Dewen;WEI Botao(School of Control and Computer Engineering,North China Electric Power University,Baoding 071000,China)
出处
《智能系统学报》
CSCD
北大核心
2021年第2期254-262,共9页
CAAI Transactions on Intelligent Systems
基金
国家自然科学基金项目(51677072)。
关键词
小样本
变分自编码器
孪生网络
图像识别
过拟合
特征向量
深度学习
数据增强
small sample
variational auto-encoder(VAE)
siamese network
image recognition
over-fitting
eigenvector
deep learning
data augmentation
