Research Of Multi-Pose Face Recognition Algorithm Based On Deep Learning

Face recognition plays an important role in our lives,it's a very popular branch of the biometric technologies,which attracts most of the researchers' wide attention and gets a great deal of achievements.However,while being used in real life,there are many unsolved problems in face recognition technology,such as in real-time monitoring system,the pthotos collected by it are often contains different postures,which will reduce the performance of the technology due to the influence of the varies postures,and there still have no effective way to fully solve the problem.The proposal of deep learning provides a new way to solve the multi-pose problem in face recognition.To solve the multi-pose problems,this paper proposes several multi-pose face recognition algorithms,and the main contents are as follows:(1)In the light of the traditional deep convolution neural networks are very difficult to convergence and are time consuming,this paper optimizes the components of the deep convolution neural network one by one.First of all,the pooling layer is replaced by the convolution structure in this paper,so that the network structure is consistent and the calculation is reduced.After that,in order to avoid the generalization of the network,the MFM+PReLU structure is used to replace the PRe LU function in the low layer of the network.At last,the momentum method is used to improve and optimize the joint supervised loss function of Center Loss and Softmax Loss,and the convergence rate is improved.We show that our model accelerates the convergence of the network and enhances the robustness of the multi-pose,the training time is shortened three hours and it is able to achieve an accuracy of 99.19% and 94.43% on both the LFW and the YTF datasets.(2)In view of the problem that the fully connected layers usually consume a lot of parameters in the current convolution neural network,convolution layer and the global average pool layer are used to replace it,which will reduce the consumption of parameters in the network training.In addition,we choose different output features in the same network to optimize feature fusion,which can simplify the training process and enhance the diversity of features.At last but not least,we use the PCA to reduce the feature dimension and study the influence of different dimensions.Our methods achieve 99.26% and 94.46% on the LFW and the YTF datasets and the training time is shorter.(3)For the insufficient of training samples in the deep convolution neural network will make the network easy to saturation and the algorithms of the current 3D model to generate virtual faces has problem to detect face images with large pose change,we combine MTCNN and SDM algorithm to get the feature points that needed for 3D mapping,and realizes the purpose of the detection of large pose varies images which will expand multi-pose samples.It reaches an accuracy of 99.33% and 95.12% on the LFW and the YTF datasets,and in the more complex multi-pose face database IJB-A,the face recognition accuracy is about 89.6%.