Research On Face Recognition Based On Deep Learning

Identity technology plays an important role in modern social life and is closely related to people's lives.Face recognition technology is intuitive,non-contact and concealed,which is widely used in security,public security,and payment.The Convolutional Neural Network(CNN),which belongs to deep learning,promotes the development of face recognition technology.Face recognition based on convolutional neural networks now has a greater advantage than traditional methods.Based on the traditional face recognition technology of artificial design features,the workload of manually acquiring features is very large,and the recognition effect is susceptible to subjective and scene limitations,resulting in low real-time and accuracy.The use of convolutional neural networks to identify faces can avoid complex feature extraction,directly learn the best image features,and be robust and real-time in different scenarios.The ResNet and DenseNet networks were recently proposed,and the model works best.ResNet can reach 152 layers,but the deep network leads to low training efficiency;DenseNet's dense connection can lead to implicit deep supervision,which helps to train deeper network architectures,but this feature reuse will lead to the introduction of redundancy in subsequent layers.By introducing the DenseNet dense connection structure with reference to the ResNet network structure,this paper designs a 12-layer CNN model architecture.Under the guarantee of certain feature reuse,but not to introduce too much redundant information,the network has good information flow transferability and implicit deep supervision.The batch normalization and Dropout improvement network are introduced in the network to further alleviate the over-fitting.The optimization algorithm for training network parameters selects Adam and the activation function selects ReLU.Ensure a certain recognition speed and accuracy,reduce the dependence on computer hardware resources,and realize online recognition of faces.The CNN model designed in the CASIA-webface face database training,and the model hyperparameters are adjusted for the training results to obtain the optimal model.By comparing with the VGG16 and ResNet14 models,the performance of the CNN model is verified to be excellent,and the test accuracy of 95.6% or more can be achieved on the CASIA-webface.The Haar Cascade classifier that comes with the OpenCV library is called to implement the face detection function,and a step of deblurring the image is added between the face detection and the face recognition,and the picture with low definition is filtered to further improve the recognition effect.Finally,the face recognition system test results in this paper are about 92.5%,and the recognition speed is within 1s,which satisfies the design requirements of the face recognition system.Figure[65] table[7] reference[70]...