A Face Recognition System Based on Convolution Neural Network

This paper presents a face recognition system based on convolution neural network. The system consists of four convolution layers, three pooling layers, one full-connected layer and one softmax regression layer. By combing the tanh activation function and the ReLU activation function together, a new activation function is proposed. Four networks with different numbers of convolution kernels in the same convolutional layers are designed by using the Theano framework in python language. Each network is trained with tanh activation function, ReLU activation function and the new activation function respectively. The ORL face database is expanded and used to train and test the networks. The results show that as the number of convolution kernels increases, the rate of misrecognition decreases and the new activation function achieves the highest performance than the other two activation functions. The system designed in this paper is efficient for face recognition.