Research On Cross-age Face Recognition Based On Deep Neural Network

As an extension of face recognition,cross-age face recognition has a wide range of practical significance.In recent years,with the development of deep learning,face recognition has become more and more mature,and has been widely used.However,the development of cross-age face recognition is slow due to insufficient databases.The first reason is that the database itself is difficult to collect.The second reason is that the existing methods do not fully use the database.Most of the methods do not use the age tags of pictures,which results in the waste of some known information.This paper focuses on cross-age face recognition,and research the combination of generation method and cross-age face recognition,the establishment of feature dictionary and how to introduce age information into network training.(1)Aiming at the problem of inadequate images in test set,a generation algorithm with identity constraints is introduced to generate missing age pictures.The recognition accuracy of FGNET database is improved by 5.5% through image feature fusion,and the accuracy of Leave-One-Out test achieves 88.0%.Through the generative method,using the age label of the test picture,ten new pictures of different age groups are generated for each test picture.Then the generated pictures are fused by features,which are divided into average fusion and weight fusion.(2)Aiming at the problem of age information in deep features,a general method of injecting age label into training network is proposed,which can help deep features eliminate age information.It can improve the accuracy by 0.4% on Morph test set and 5.4%when the FAR equals 0.001.In this method,the age label is firstly transformed into binary coding.Then,it is nonlinearized and scaled through several fullly connection layers.Finally,it cascades with the features extracted from the deep network and continues to be sent to calculate the loss.(3)Aiming at the application of age labels in loss function,a loss function with age label is proposed,which makes the loss of images(too old or too young)larger.The recognition accuracy can be improved by 0.7% on Morph dataset and 3.2% when FAR is limited to 0.001.In this method,age label is firstly converted into age coefficient and then multiplied with the softmax loss or center loss of the image to obtain a new loss.The sum of all image losses is the total loss function.