Face Verification Based On Deep Neural Networks

As a non-contact authentication method,face verification has unique advantages compared with fingerprint recognition,and is suitable for many scenarios.Recent years,the development of deep learning has greatly promoted the performance of face recognition,making the development of face recognition reach an unprecedented height,and many projects are gradually falling.Based on the realistic needs of face verification in many realms in real life,this thesis studies the topic of face verification based on deep neural networks.First of all,this thesis elaborates on the background and the significance of the project,and investigates the current research status of the topic at home and abroad,and determines the significance and feasibility of the research topic of this thesis.Secondly,this paper investigates three main factors affecting the performance of face recognition based on deep neural networks:data,network structure,loss function,and verifies their influence through experiments.In this process,this paper transferred the model trained by Western faces to adapt to oriental faces through transfer learning,and achieved good results.Moreover,in the process of training the face recognition model,a data cleaning scheme is proposed for the non-convergence encountered.After data cleaning,a model is successfully trained.In addition,the data augmentation is used to improve the generalization ability of the model.On this basis,the whole process of face recognition is sorted out,and different feature extraction methods are compared and applied to the face recognition system.Then,this thesis conducts a face-based study on the common attributes:gender and age.According to the characteristics of age recognition,the influence of the adjacent interval is introduced into the loss function,and good results are obtained.Finally,the overall research content is summarized,and the future work is forecasted for the existing problems.