Research On The Cross-race Face Recognition Algorithm Based On Deep Domain Adaptation

Face recognition is a popular and mature biometric recognition technology in the field of computer vision research.It has a place in video security,financial authentication and shopping payment.With the development of deep learning,the huge amount of data and large-scale computing capabilities have continuously refreshed the baseline of face recognition and even exceeded the human level.However,in complex unconstrained scenes,many challenges are faced.Changes in light intensity,pose range,and face attributes greatly increase the difficulty of the recognition process.Compared with the constraints under a single closed set,the identities of the training and testing set in the open set task of unconstrained scenarios do not contain any intersection,so the difference in the data distribution between the training and testing set affects the final identification to a certain extent.Most of the public data sets are composed of white people.If the model trained with the regular training set is used in our daily life,that is,it will be used on the yellow testing set,which will lead to an extreme decrease in accuracy.Massive samples can be easily obtained through web crawlers,but at the same time the cost of obtaining a large number of labeled yellow samples is time-consuming and labor-intensive.Therefore,how to use the limited data to eliminate the distribution difference between different races is particularly important.Although the obtained samples require complex and tedious manual labeling to remove and clean them,they can be solved by semi-supervised learning through a large number of labeled Caucasian samples combined with a small number of unlabeled yellow samples.In view of the above deficiencies,this paper uses the domain adaptation method in transfer learning to propose an improved deep domain adaptation network framework that is suitable for face recognition tasks.The research work of this paper is as follows:(1)To solve the problem that the learned features migrates from the current domain to the target domain can improve the recognition effect of the target domain,meanwhile,the accuracy in the source domain is not guaranteed when using the common domain adaptation method.This paper learns the overall distribution of the data through batch normalization.At the same time,the style invariance of instance normalization is learned.In order to reduce the difference between the source domain and the target domain,a dynamic balance factor is introduced,and then a domain adaptation method Ada-IBN based on dynamic adjustment and normalization is proposed,and a loss function based on angular interval is combined to improve discrimination of features.(2)The existing domain adaptation method can only reduce the domain difference during the process,but not completely eliminate the domain difference,and the negative migration generated during the domain adaptation process cannot be avoided.This paper proposes a cross-racial face recognition network with domain space alignment and learning discriminative features.First,pseudo-label samples are generated by clustering algorithm to improve the recognition effect of the target domain.On the other hand,most of the existing domain adaptation methods map different features to the same subspace for alignment,but they neglect to eliminate the negative transfer problem caused by cross-domain.On this basis,this paper introduces the Sliced Wasserstein distance that measures different distributions as the penalty term reduces negative migration,while also increasing the ability to discriminate features.(3)In order to have a deeper understanding of study the impact of cross-ethnicity on face recognition tasks,in addition to the public standard testing set and the blank of the existing yellow testing set,this paper has made an Asian face test set containing thousands of identities set named Cceleb-1K.Based on this comparison,experiments are conducted to verify the effectiveness and superiority of the proposed algorithm.