基于卷积神经网络的人脸表情识别研究
|
摘要
随着人机交互技术和机器学习技术的发展,人脸表情识别技术逐渐成为研究热点。针对传统人脸表情识别算法鲁棒性差、表情特征提取能力不足的问题,提出一种改进的基于卷积神经网络的人脸表情识别算法。首先对人脸图像进行预处理,检测并分割出人脸关键点的部分图像,然后输入到包含卷积神经网络通道和卷积稀疏自编码(CSAE)预训练通道的双通道模型中。其中卷积神经网络通道部分使用了批量正则化(Batch Normalization)和ReLU激活函数,加快了模型训练速度,解决了梯度消失问题,同时增加了模型的非线性表达能力。通过引入Dropout技术,解决了网络的过拟合问题。在另一个通道,对输入的人脸表情图像增加了卷积稀疏自编码进行无监督预处理。实验结果表明,该算法在JAFFE、CK+人脸表情数据集上均获得了较好的识别效果。
With the development of human-computer interaction technology and machine learning technology,facial expression recognition technology has gradually become an important field.In this paper,we proposean improved algorithm based on Convolutional Neural Network(CNN)for face expression recognition due to the lack of robustness of traditional facial expression recognition algorithm and availabe feature extraction ability.First of all,we pre-train the facial image and detect,segment face.The segmented face applied to a dual-channel model which,includes a convolutional neural network channel and an extra pretraining channel by sparse convolutional autoencoders.The training speed is improved according to Batch Normalization and the ReLU activation function by convolutional neural network channel,and solve the problem of gradient disappearance.This modelcan increase non-linear expression ability of the model.At the same time,the introduction of dropout technology also remove the problem of over-fitting.Another channel that contains a sparse convolutional autoencoder aims to deal with input facial expression images.Experimental results involved this algorithm demonstratean improved recognition ability on the JAFFE and CK+dataset.
With the development of human-computer interaction technology and machine learning technology,facial expression recognition technology has gradually become an important field.In this paper,we proposean improved algorithm based on Convolutional Neural Network(CNN)for face expression recognition due to the lack of robustness of traditional facial expression recognition algorithm and availabe feature extraction ability.First of all,we pre-train the facial image and detect,segment face.The segmented face applied to a dual-channel model which,includes a convolutional neural network channel and an extra pretraining channel by sparse convolutional autoencoders.The training speed is improved according to Batch Normalization and the ReLU activation function by convolutional neural network channel,and solve the problem of gradient disappearance.This modelcan increase non-linear expression ability of the model.At the same time,the introduction of dropout technology also remove the problem of over-fitting.Another channel that contains a sparse convolutional autoencoder aims to deal with input facial expression images.Experimental results involved this algorithm demonstratean improved recognition ability on the JAFFE and CK+dataset.
引文
[1]牛玉虎.卷积稀疏自编码神经网络[J].计算机与现代化,2017(2):22-29.
[2]李江,冉君军,张克非.一种基于降噪自编码器的人脸表情识别方法[J].计算机应用研究,2016,33(12):3843-3846.
[3]MAKHZANI A,FREY B.A winner-take-all method for training sparse convolutional autoencoders[J].Eprint Arxiv,2014.
[4]VU T D,YANG H J,NGUYEN V Q,et al.Multimodal learning using convolution neural network and sparse autoencoder[C].IEEE International Conference on Big Data and Smart Computing,2017:309-312.
[5]RIFAI S,VINCENT P,MULLER X,et al.Contractive auto-encoders:explicit invariance during feature extraction[C].ICML,2011.
[6]BURKERT P,TRIER F,AFZAL M Z,et al.DeXpression:deep convolutional neural network for expression recognition[J].Computer Vision and Pattern Recognition,2015,22(10):217-222.
[7]CHAI R.Face recognition algorithm based on Gabor wavelet and deep belief networks[J].Journal of Computer Applications,2014(9):1938-1943.
[8]MASCI J,MEIER U,DAN C,et al.Stacked convolutional autoencoders for hierarchical feature extraction[M].Artificial Neural Networks and Machine Learning–ICANN 2011.Springer Berlin Heidelberg,2011:52-59.
[9]HINTON G E,SALAKHUTDINOV R R.Reducing the dimensionality of data with neural networks[J].Science,2006.
[10]DOI E,INUI T,LEE T W,et al.Spatiochromatic receptive field properties derived from information-theoretic analyses of cone mosaic responses to natural scenes[J].Neural Computation,2003,15(2):397-417.
[11]LOPES A T,AGUIAR E D,OLIVEIRASANTOS T.A facial expression recognition system using convolutional networks[C].Graphics,Patterns and Images.IEEE,2015:273-280.
[12]IOFFE S,SZEGEDY C.Batch normalization:accelerating deep network training by reducing internal covariate shift[J].Learning,2015:448-456.
[13]HINTON G E,SRIVASTAVA N,KRIZHEVSKY A,et al.Improving neural networks by preventing co-adaptation of feature detectors[J].Computer Science,2012:212-223.
[2]李江,冉君军,张克非.一种基于降噪自编码器的人脸表情识别方法[J].计算机应用研究,2016,33(12):3843-3846.
[3]MAKHZANI A,FREY B.A winner-take-all method for training sparse convolutional autoencoders[J].Eprint Arxiv,2014.
[4]VU T D,YANG H J,NGUYEN V Q,et al.Multimodal learning using convolution neural network and sparse autoencoder[C].IEEE International Conference on Big Data and Smart Computing,2017:309-312.
[5]RIFAI S,VINCENT P,MULLER X,et al.Contractive auto-encoders:explicit invariance during feature extraction[C].ICML,2011.
[6]BURKERT P,TRIER F,AFZAL M Z,et al.DeXpression:deep convolutional neural network for expression recognition[J].Computer Vision and Pattern Recognition,2015,22(10):217-222.
[7]CHAI R.Face recognition algorithm based on Gabor wavelet and deep belief networks[J].Journal of Computer Applications,2014(9):1938-1943.
[8]MASCI J,MEIER U,DAN C,et al.Stacked convolutional autoencoders for hierarchical feature extraction[M].Artificial Neural Networks and Machine Learning–ICANN 2011.Springer Berlin Heidelberg,2011:52-59.
[9]HINTON G E,SALAKHUTDINOV R R.Reducing the dimensionality of data with neural networks[J].Science,2006.
[10]DOI E,INUI T,LEE T W,et al.Spatiochromatic receptive field properties derived from information-theoretic analyses of cone mosaic responses to natural scenes[J].Neural Computation,2003,15(2):397-417.
[11]LOPES A T,AGUIAR E D,OLIVEIRASANTOS T.A facial expression recognition system using convolutional networks[C].Graphics,Patterns and Images.IEEE,2015:273-280.
[12]IOFFE S,SZEGEDY C.Batch normalization:accelerating deep network training by reducing internal covariate shift[J].Learning,2015:448-456.
[13]HINTON G E,SRIVASTAVA N,KRIZHEVSKY A,et al.Improving neural networks by preventing co-adaptation of feature detectors[J].Computer Science,2012:212-223.