Research On Face Recognition Algorithm Based On Deep Learning

As a hot topic in academics and industry,face recognition has been widely used in intelligent safeguard systems,financial security,mobi-le payment and other fields.However,with the rapid expansion of the business scope,the number of identities in the face database has also increased steeply,and the false alarm rate of the current algorithm obviously cannot meet the growing application requirements.Therefore,reducing the false alarm rate in large-scale face recognition scenes is becoming a hot issue in face recognition research.The most direct and effective method is to extract more distinguishable features.Recent face recognition algorithms have been able to achieve better recognition capabilities by adding margin to ground truth,but to further improve the recognition performance,it is also necessary to make full use of the information of the negative classes and further dig the hard samples in them.Based on the analysis above,on the one hand,this paper focuses on the loss function and mine the information of the negative classes in the training dataset;on the other hand,this paper pays more attention to the sampling strategy and help the model achieve more efficient learning by introducing similar samples during sampling.Specifically,the main contributions of this paper are reflected in the following three aspects:First,this paper proposes a new loss function for hard examples mining in face recognition.It divides the negative classes into simple negative classes and hard ones by recognizing the hard samples in the negative classes.And it keep the model focused on the hard samples to achieve more targeted optimization.Second,this paper jointly optimizes the doppelganger mining and the proposed loss function.The two methods promote each other and achieve more efficient and accurate hard example mining.This paper draws on the ideas of curriculum learning to further improve the proposed loss function.By changing fixed hyperparameters to parameters that change continuously with the training process,adaptive mining of hard samples in the negative classes is realized.It helps the loss function match the training process,which not only helps the model to converge,but also improves the ability of hard example mining.