Research Of Face Recognition Based On Convolutional Neural Network

The development of computers is changing rapidly,by the biometric identification of personal information has become an ideal and reliable means of verification,face recognition on facial features is favored by the academia and industry by virtue of the features of directness,rapidity,non-contact,etc.But in the actual scene,face recognition often by much interference of light,angle,occlusion and so on.And deep learning can approach to complex functions by learning a deep nonlinear network structure,it has a powerful ability of feature extraction of the input sample data simultaneously.The convolution neural network has become a hotspot currently in the field of face recognition because of its three main ideas:local connection,weight sharing and pooling sampling.This paper fully investigates the significance and background of this topic,and analyzes the main difficulties of face detection and recognition through the current situation at home and abroad,and makes some research on the following aspects:(1)Face detection based on MTCNN.MTCNN consists of three network cascades,each of which is responsible for processing multiple tasks,including detection tasks,alignment tasks,etc,and uses an effective online difficult sample mining method to further enhance the performance of the network,not only is suitable for the detection of multiple target face,and real-time performance and good robustness.Through extensive experiments on public data sets and real data sets,the results show that MTCNN has powerful performance and high detection rate.(2)Face recognition based on Inception-ResNet-V1.Inception-ResNet-V1 network combines the idea of residual learning,which ensures its excellent ability of face image representation and avoids the problem of the accuracy degradation caused by the deepening of network level.Similarly,that network is also verified in both public data sets and real data sets,experiments show that trained Inception-ResNet-V1 model under the condition of meet the real-time requirements,also has a good identification accuracy and strong adaptability for profile and occlusion.(3)Applying the above research to real scene,a multi-person attendance system based on real-time monitoring is designed.The system includes four modules:video stream analysis,face detection,face recognition and log record.According to the actual situation,this thesis puts forward apply the frame differential method to video analysis,on the basis of the combination of face region area selection,to avoid a lot of fuzzy,poor illumination image,but also improve the efficiency of the system.Experiments show that the accuracy of face detection and recognition is improved by the above optimization.The test results in the dataset of the real scene have reached 98.25%,which has a strong practical value.In summary,this paper adopts MTCNN and Inception-ResNet-V1 network for face detection and recognition,and carried out the experiment in public data sets and real data sets,have achieved higher accuracy,so the research work of this paper has a certain theoretical significance and application value.