Research On Face Recognition Method For Improving VGG Network

Face recognition,as a branch of biometrics,is widely used in many fields such as security monitoring and mobile payment.With the rise of deep learning,various face recognition models based on convolutional neural network(CNN)have emerged endlessly,but they are generally divided into two steps: face detection and recognition.The thesis is optimized on the basis of the classical face recognition network VGGFACE(visual geometry group-face)and the detection network MTCNN(multi task convolutional neural network).The following aspects are studied:1)The problem of insufficient diversity of the face data set is studied.The current mainstream face recognition databases generally use Westerners as a sample.The lack of diversity in the database will cause the face recognition model to have a low accuracy rate for face recognition of a specific skin color and race.In response to this problem,the paper obtained some photos of Chinese stars from the Internet,and combined the face images collected by the laboratory classmates to enhance the data,so as to supplement the database.2)The problem of face detection in the case of small face spacing is studied and a BMTCNN model is proposed for this problem.MTCNN takes into account both the speed and accuracy of detection,but the network's detection effect decreases significantly when the face spacing is too small.To solve this problem,R-Net and O-Net in MTCNN are improved.The NMS algorithm of each network module is optimized as a better-NMS algorithm,which predicts the Io U(intersection and union ratio)of the target candidate bounding box and the real target label,and uses the product of this and the classification score as the detection confidence.The improved model is BMTCNN.The experimental results show that BMTCNN has a good detection accuracy in the case of small face distances,and the missed detection rate has decreased by 5.9% compared with MTCNN.3)The problem of reducing model parameters while ensuring recognition accuracy is studied,and two models of DVGGFACE and BVGGFACE are proposed.VGGFACE is a 19-layer network but contains parameters.A large number of parameters are not conducive to the model's ability to resist overfitting,and the recognition accuracy of VGGFACE is not dominant compared to the current model.In view of the above problems,the use of global average pooling to replace the fully connected layer in VGGFACE has greatly reduced the amount of parameters;the use of dense connection technology to improve VGGFACE to obtain DVGGFACE,that is,the same convolution layer in the convolution module in VGGFACE is fast The connection allows each layer in the network to accept the features of all the layers in front of it as input,thereby improving the recognition accuracy of the model;optimizing part of the convolutional layers in DVGGFACE to deeply separable convolution to obtain BVGGFACE.The experimental results show that the optimized DVGGFACE improves the recognition accuracy(compared with VGGFACE by 4.67%)and significantly reduces the parameters(compared with VGGFACE by 71.61%,and the parameter amount is reduced by an order of magnitude).The paper finally achieved higher recognition accuracy with fewer parameters,laying the foundation for the model to be ported to the mobile end.