Design And Research Of A Face Recognition System Based On Deep Learning

In recent years,due to the more powerful feature learning and feature expression capabilities,convolutional neural networks based on deep learning algorithm training have made remarkable achievements in the field of computer vision.In view of this,based on the deep learning algorithm,using Tensor Flow deep learning framework and Python language,this paper studies the real-time facial identity and facial expression recognition in video frames under the left and right tilt conditions.main tasks as follows:(1)The structure of the traditional Le Net-5 network has been improved.In the improved Le Net-5 network structure,a 3*3 convolution kernel with fewer parameters and a Softmax classifier with strong nonlinear classification ability are used;and inspired by the Inception network,convolution kernels of different sizes are used for the input layer of the Le Net-5 network.To extract features.Furthermore,the validity of the network structure is improved by experimental verification on the orl face dataset.By controlling the same variable method,the different optimization algorithms of Re LU,Tanh and Swish are compared between BGD,Adagrad,Adadelta and Adam.,the difference in the impact on network performance,in preparation for the selection of subsequent experimental activation functions and optimization algorithms.(2)In order to improve the structure of the VGG-16 network,the global average pooling technique is used to reduce most of the training parameters of the network by removing all the fully connected layers;the batch normalization technology is used after the partial convolutional layer to optimize the input of the layer.The distribution of data is more conducive to network training.And test the effectiveness of the improved network structure on the cifar-10 data set.(3)Combined with MySQL database,the face recognition system is designed under the condition of left and right face tilt.In order to achieve clear and conducive effect on network recognition,the histogram of the input network is subjected to histogram equalization and median filtering;and the Inception network is used to extract features and VGG-16 network double volumes in different convolutional layers.The advantages of the core stacking extraction feature,the global average pooling and batch normalization techniques are added,and the face identity model CNN-1 and the face expression model CNN-2 are designed and trained on the relevant data sets.Finally,the entire face recognition system was tested comprehensively.Tests show that the designed face recognition system has good recognition performance.