Research Of Face Recognition Based On Convolutional Neural Network

Convolutional neural network is a multi-layer perceptron, which is specially designed to identify the two-dimensional shape. Convolutional neural network is trained by back-propagation and gradient descent algorithm. Convolutional neural network integrates the ideas of local receptive fileds, shared weights and downsampling into its structural design, which makes it highly invariant to translation, scaling, tilting or other forms of deformation to a certain degree. Currently, convolutional neural network has been widely used in pattern recognition and image processing applications.In this paper, we intensively investigate the theory of the artificial neural network and the convolutional neural network, and construct some convolutional neural network models on the basis of the famous LeNet-5 structure. The highlights of our work include:(1) Face alignment is important to face recognition and it depends on the eye location, so in this paper, we propose a cascaded convolutional neural network model for eye detection under complex scenarios. Compared with other methods, the detection performance of our cascaded convolutional neural network is much better than those of other methods.(2) We investigate the performance of three convolutional neural networks with different network complexities, on the Yale B face datebase, PUT face database and AR face database. Firstly, we combine different activation functions with different downsampling methods and then analyze the recognition ability of different combinations on the three face databases, respectively. Secondly, we improve the recognition performance from three aspects respectively:adding weight decay term in the cost function; applying the dropout technology in the network; adding a momentum term in the weight updating formula. Thirdly, we fuse the above three aspects into the network, and compare its performance with the network without the three aspects. Additionally, by removing the output layer of convolutional neural network and feeding the fully connected layer into SVM classifier, we compared the features extracted by the network with and without three terms. Experimental results show that the performance of the network with three terms is better than the network without three terms.(3) We apply the two-dimensional Gabor filter in the structure of convolutional neural network(GCNN) and construct 5 GCNNs from 5 scales respectively. We compare and analyze the performance of three methods:GCNN, GCNN+SVM(GCNN extracts the features and then feeds the features into SVM classifier), PCA+SVM(PCA reduces the dimensionality of the Gabor features and then feeds the features into SVM classifier). In addition, combining the 5 GCNNs into a multi-column GCNN(MGCNN), we compare its performance with PCA+SVM method and evaluate the impact of the dimensionality change of feature on both methods. Experimental results showed that MGCNN outperforms PCA+SVM, especially with very low-dimensional features.