Research On Face Recognition Technology Based On Convolutional Neural Network

With the development of artificial intelligence technology,biometric recognition has become a new-type human-computer interaction for many industries.Among them,face recognition technology has gradually become one of the most important recognition methods because of its non-contact,non-invasive,simple equipment and difficulty to crack it.Convolutional neural network is widely used in face recognition technology.It usually uses the cosine domain loss function to expand the margin between features,which will increase the training time.Therefore,in my dissertation,I have devoted my time to research on network training speed and the actual recognition effect in the cosine domain.First of all,for the time-consuming problem of network training in the cosine domain,this paper proposes a cosine margin loss function with a variable slope.This function adds a slope factor to the cosine margin loss,so that the intra-class constraints gradually increase with the increase of the cosine value,and the intra-class distance is explicitly reduced while taking into account the speed.1:1 verification on LFW and Agedb shows that compared with other cosine domain loss functions,neural networks training with the variable slope cosine domain loss can reduce the time consumption by more than 15.2%with the same recognition rate.Secondly,in order to optimize the recognition effect,aiming at the shortcoming of Euclidean distance normalized by L₂ norm to measure feature similarity,this paper proposes a method combining feature distance and binary classification.Among them,the feature distance is a multiple-distance extended on the basis of the Euclidean distance.Experiments show that this method can improve the face recognition rate by about 0.17%in less than 2%space.At last,based on the above models and matching methods,this paper designs an access control system based on face recognition technology.The system has the functions of monitoring and photo attack prevention,and its face recognition module and human-computer interaction module can run across platforms.Experiments show that the system's recognition speed is0.8s/frame and error-traffic rate is less than 0.5%.The operating conditions meet the requirements of GB/T 35678-2017.