基于变分自编码器的人脸表情识别

FACIAL EXPRESSION RECOGNITION BASED ON VARIATIONAL AUTOENCODER

为了在样本量较小的表情数据集上实现较高准确率的表情识别,将无监督特征学习应用于表情识别中,在传统自编码器网络的基础上,将变分自编码器引入人脸表情识别中,提出一种基于变分自编码器改进的人脸表情识别方法,使用大量无表情标签的人脸数据集对变分自编码器进行无监督训练,将变分自编码器中编码网络部分输出的低维特征输入到卷积神经网络中,由变分自编码器的编码网络和卷积神经网络两部分构成完整的表情识别网络;使用带表情标签的人脸表情数据集对网络进行训练。在CK+、JAFFE数据库进行分类实验,实验结果表明,该算法具有一定的表情识别能力。

In order to achieve high accuracy expression recognition on expression data set with small sample size,unsupervised feature learning is applied to expression recognition.On the basis of traditional self encoder network,variational autoencoder was introduced into facial expression recognition.A facial expression recognition method based on improved variational autoencoder was proposed.A large number of facial data sets without expression labels were used to train the variational autoencoder unsupervised.The low-dimensional features which were output from the encoding network part of the variational autoencoder were put into the convolutional neural network,and a complete expression recognition network was composed of the encoding network and the convolutional neural network.The facial expression dataset with expression tags was used to train the network.The classification experiments were conducted on CK+and JAFFE databases.The results show that this algorithm has a certain ability of expression recognition.