Deep Learning-based Age-Invariant Face Recognition Method

As a branch of face recognition(FR),age-invariant face recognition(AIFR)has significant research value and is widely applied in the world of finance,security,smartphone applications,and so on.However,affected by the deeper facial wrinkles and changes in skull shape over time,AIFR is usually more difficult and less accurate than common FR.What's more,the research of AIFR starts later than common FR,leading to less literature on it.For the above reasons,this paper carries out some research on the techniques of AIFR as follows.Firstly,with the help of the ResNet50 network which performs well in common images feature extraction,we introduce the deformable convolution and the dilate convolution and propose an end-to-end facial feature extraction model based on it.Particularly,we combine the feature extraction and the cosine similarity to perform a better end-to-end AIFR method.Extensive experiments demonstrate a satisfactory AIFR accuracy of the given method.Secondly,considering that the measurement with Siamese network is helpful to improve the performance of intra-class aggregation and interclass separation,we combine the siamese network architecture and the feature extraction model mentioned above to further improve the feature extraction ability of the model,And then,the siamese AIFR method is proposed with the integration of the cosine similarity.Moreover,a measurement with multivariate siamese relationship is proposed and incorporated into the modified model novel.Extensive experiments show that the method based on multivariate siamese relationship achieve better performance on AIFR.