Design And Implementation Of Face Recognition System Based On Convolutional Neural Network

Face recognition is a kind of bio-technology based on the information of human facial feature for identity recognition.It has a wide range of applications in identity authentication scenarios,certificate verification scenarios and face retrieval scenarios due to its intuitive,reliable and stable advantages.Many face recognition algorithms have been proposed by scholars,but there are still two problems in current face recognition algorithms: on the one hand,the recognition rate of face recognition algorithms on some open data sets is still low;on the other hand,most of the current mainstream face recognition algorithms use convolutional neural network as feature extraction model.The use of convolution neural network will result in the long training time of face recognition algorithm.In this paper,a large number of papers and reports on convolutional neural networks and face recognition are fully investigated,the main work of this paper can be listed as following:(1)Through the analysis of feature extraction,activation function and loss function in Sphere Face 20,the original network is optimized by SE-Res Net module and Swish activation function,and the face recognition network SE-Res Net-20 is obtained.In SE-Res Net-20 network,the use of SERes Net module improves the effect of face image feature extraction,besides,Swish activation function improves the non-linear fitting ability of the network.Finally,after training with CASIA-Webface data set,the recognition rates of SE-Res Net-20 network on LFW data set and Mega Face data set are 99.37% and 68.92%,respectively,which are 0.15% and 5.49% higher than those of Sphereface20 network.(2)Aiming at the problem of long training time of convolutional neural network,this paper optimizes the training of SE-Res Net-20 network by Caffe On Spark cluster method.On this basis,combined with asynchronous training and momentum weight updating method,the Caffe On Spark training method of asynchronous momentum gradient descent is obtained and the training of SERes Net-20 network is further optimized.The experimental comparison shows that the training time of SE-Res Net-20 network can be reduced by nearly 35% by using Cafe On Spark training method with asynchronous momentum gradient descent.(3)By analyzing the flow of face recognition algorithm,the paper designs and implements a set of face recognition system.According to the actual functional requirements,the system is divided into five modules: image acquisition module,face detection module,pre-processing module,face recognition module and system management module.With the cooperation of these modules,the face recognition system can complete face image recognition,face image comparison,face image acquisition and face data storage effectively.Through various tests,the whole system can work steadily and fulfill the predetermined functional requirements.