Reasearch Of NIR-VIS Heterogenous Face Recognition Methods Based On Facial Attributes

In this paper,research is conducted on NIR-VIS heterogenous face recognition methods based on facial attributes.NIR and VIS images are formed by near infrared light and visible light.Different imaging mechanisms cause huge gap between these two kinds of images.NIR-VIS heterogenous face recognition refers to recognition based on computation of NIR and VIS images,while in conventional homogenous VIS-VIS face recognition,images all come from the VIS domain.Therefore,NIR-VIS heterogenous face recognition differs from VIS-VIS homogenous face recognition in that algorithms need to remove the heterogenous effects on face recognition besides the conventional interference factors like expression,pose and illumination strength.To improve the performance of NIR-VIS heterogenous face recognition,we propose a NIR-VIS heterogenous face recognition method based on facial attributes.This method combines facial attribute prediction and heterogenous face recognition through multi-task learning,aiming to boost the heterogenous recognition performance via facial attribute prediction task.Meanwhile,in face classification and cross domain learning modules,this method utilizes softmax loss and triplet loss to learn heterogenous invariant feature representations.Main work of this paper is divided into four parts: proposing a unified NIR-VIS heterogenous face recognition framework,facial attribute extraction of VIS and NIR face datasets,multi-task learning of facial attribute prediction and face recognition in VIS domain,heterogenous face recognition based on facial attributes.A unified NIR-VIS heterogenous face recognition framework is concluded in this paper.Mainstream NIR-VIS heterogenous face recognition methods can be unified by this framework.It is found that cross domain learning is a key part in NIR-VIS heterogenous face recognition.NIR-VIS heterogenous face recognition method based on facial attributes is proposed based upon this framework.In attribute extraction of NIR and VIS datasets section,to improve the accuracy of facial attribute prediction,voting strategy is adopted to determine the attributes of each person.In VIS-domain face recognition and attribute prediction section,different ways of multi-task learning are analyzed to determine the best way of learning.Contrast analyses of multi-task and single-task algorithms prove that facial attribute prediction task has a positive effect on face recognition.In NIR-VIS face recognition section,the face classification and cross domain learning modules are carefully designed.Softmax loss and triplet loss are utilized to learn the heterogenous invariant feature representations.Theoretical and experimental analyses are made about the importance of feature normalization and scaling to triplet loss.Through contrast experiments,the best learning strategy of attribute prediction task in multi-task learning is determined.Final evaluation results show that on rank-1 criterion,the proposed method is only a little lower than CDL algorithm but beats other mainstream algorithms.However,on verification criterion where the false accept rate is 0.1%,the proposed method surpasses all the current algorithms.