Research on Facial Recognition based on Convolutional Neural Network

  • Guoping Lei*, Xiuying Luo , Quntiao Li , Ke Xiao , Li Deng, Minlu Dai

Abstract

The performance of the three activation functions was compared aiming at the problems such as
excessive model training time and failure of rapid network convergence. ReLU, Tanh and
Sigmoid, based on the 14-layer convolutional network model. Since the function of activation
function is to introduce nonlinear elements into the neural network and make the neural network
complete the nonlinear mapping, therefore, without the activation function, the outputs are
simply a linear combination of inputs no matter how many layers of the neural network there
are. Therefore, the activation function has a very important impact on the nonlinear ability of
the network. Finally, in view of the over-fitting problem in neural network training, Dropout
technology is introduced in the network model of the 14-layer convolutional neural network,
and the influence of Dropout on network performance is compared when Dropout is 0.3, 0.5 and
0.8 respectively. The experiment result shows that the ReLU function has the best network
performance among the three activation functions under the 14-layer model. After the
introduction of Dropout technology, the network is sparser, the generalization ability of the
model is higher, and the model is more abstract. Among the three different parameters of
Dropout, the network performance is the best when the parameter is 0.5. By applying this model
to the conference face check-in system, the results are good at present.

Published
2020-06-30
How to Cite
Guoping Lei*, Xiuying Luo , Quntiao Li , Ke Xiao , Li Deng, Minlu Dai. (2020). Research on Facial Recognition based on Convolutional Neural Network. Design Engineering, 206 - 221. https://doi.org/10.17762/de.vi.453
Section
Articles