Studies On Key Issues Of Unconstrained Face Recognition

As one of the most promising technologies in the field of computer vi-sion and artificial intelligence,face recognition has been widely concerned by academia and industry for many years.At present,face recognition under con-strained environment has achieved promising results,and many products with face recognition technology have entered the market.However,it is still a chal-lenging problem on unconstrained environment,especially due to factors such as the variation of pose,illumination,expression,occlusions and the loss of in-formation in the data itself,leading to a poor performance of face recognition in this scenario.Based on face 3D model and deep learning,this dissertation intensively studies several issues of face recognition under unconstrained en-vironment.Through theoretical and experiments on several unconstrained face datasets,the validity of the proposed algorithm is verified in this dissertation.The innovations and contributions are as follows:(1)In order to overcome the issue of insufficient data for training process in unconstrained face recognition based on deep learning,a face augmentation method based on 3D morphable model is proposed.The face image is aug-mented from three aspects:facial pose,expression regularization,and illumi-nation.The augmented data combined with original data can be used to train the deep network,which can not only solve the problem of difficult data collec-tion,but also enhance the intra-class diversity of data.Furthermore,in order to alleviate the bias between the augmented data and the real data,this dissertation proposes an adaptive face recognition method based on data augmentation and designs a deep transfer network(DTN).The DTN projects source domain sam-ples and target domain samples to a new space,where they are fused together such that one cannot distinguish which domain a specific image is from.This method can effectively release the problem of inconsistent data distribution.(2)Aiming at the problem of lacking labeled data in few-shot face recogni-tion,this dissertation proposes a synergistic domain adaptation network(SDAN),which uses face classification,domain adaptation,and pseudo-label data learn-ing to learn a shared encoding representation.During training,this network firstly uses the labeled data and unlabeled data to train the face classification and domain adaptation modules,and then learns face features based on these two modules.Meanwhile,the domain adaptive modules can distinguish which data are labeled data.Secondly,the network can calibrate some pseudo-label data autonomously,and these data can be used as labeled data again to train the model.In the whole training process,three modules are trained alternately and cooperatively,so that the whole model could learn more discriminant face features,so as to effectively solve the problem of few-shot face recognition.(3)Aiming at pose robustness in unconstrained face recognition,this dis-sertation proposes an adaptive pose alignment(APA)method.Instead of align-ing all faces to the predefined,uniform frontal shape,we adaptively learn the alignment templates according to the facial poses and then align each face of training or testing sets to its related template.The proposed method can greatly reduce the intra-class difference and correct the noise caused by the traditional method in the alignment process,especially in unconstrained settings.Fur-thermore,this dissertation proposes an APA-based network(APANet).The APANet does not need to perform landmarks detection.By inputting a mis-aligned face image,and the model will output an aligned face image.Experi-mental results show that the APANet can achieve almost the same results as the APA method.At last,this dissertation analyzes and compares the advantages and disad-vantages of proposed methods,and discusses some future research directions.