The Research Of Face Recognition Technology Based On Deep Learning

Face Recognition,as a popular technology in the field of computer vision,has been widely applied to intelligent security,community services,mobile payment and so on in recent years.With the rapid development of deep learning,how to extract discriminative features by Convolutional Neural Networks(CNNs)is the core task of face recognition.To solve the problem of low feature discrimination in large-scale face recognition,this thesis will optimize the face recognition algorithms from the feature embedding of CNNs and the multi-margin constraint of loss function to enhance the generalization ability of models.The main research contents include:1.Aiming at the problems of information redundancy among channels and the mechanism of sharing convolutional kernel parameters,which do not give different weights to the spatial information during the feature embedding process by standard CNNs,this thesis proposes channel attention model and spatial attention model with the introduction of attention mechanism.They two learn inter-channel and inter-spatial relationships and give different weight of channel and spatial dimension of feature maps to reduce the information redundancy among channels and focus on the most important part of face images,finally leading to an optimized face feature.2.Channel Attention Model and Spatial Attention Model exist the problem of overcompressing information and do not take global features into consideration during the process of attention computation.In this thesis,the improved self channel attention model and self spatial attention model are proposed with the help of self attention mechanism.In the process of self attention calculation,they retain more data to prevent information loss.At the same time,they optimize the features by calculating the global inter-channel and inter-spatial relationship through cross-correlation matrix.3.To solve the problem that the existing loss functions can not constrain all distributed samples well,a multi-margin based loss function is proposed in this thesis.It constrains the angle and cosine value in both the feature hypersphere space and the feature optimization space respectively,which can effectively enhance the discrimination ability of the face models.In addition,in order to solve the problem that existing multiple datasets can not be merged directly in practice,this thesis proposes a training mode of multi-task alternating learning to enhance the ability of multi-data collaboration.Experiments on three small test datasets: LFW,AgeDB-30,CFP-FP and a million scale large test dataset: MegaFace show the effectiveness and feasibility of the proposed methods.In the million scale tests,the face recognition rate finally is improved to 98.38% and the face verification rate finally is improved to 98.45%,which outperforms most of the existing state-of-the-art algorithms.