Domain Adaptive Face Representations And Face Recognition Algorithm Based On Transfer Learning

As an important subject in the field of computer vision,face recognition has been widely used in the applications of security,economy,and entertainment.The current state-of-the-art algorithms show great performance in homogeneous scenes,and the evaluation results on some public testing data sets have reached a level beyond humans.However,in actual scenes,complex background still make face recognition face several difficulties.Changes of facial expressions and attitudes,disturbances from background environment,occlusion caused by decorations,blurred images caused by moving or devices often lead to significant decrease in accuracy.It still needs more efforts to improve the performance of face recognition in open scenes.Based on transfer learning,this paper proposes an efficient domain adaptive face recognition algorithms and designs an elegant and lightweight network with great practical value for training and deployment.Through the multitask training framework,considerable samples from different domains can be well absorbed and the network can better use the specific domain information hidden in multiple task branches to improve the generalization ability of the model.At the same time,this paper proposes an adaptive parameters loss function,which can automatically adjust the hyper-parameters according to the distribution of the training samples.It helps to apply suitable constraint during the training process,thereby further improving the recognition accuracy.Besides,this paper also proposes a transfer learning method to enhance the learned domain invariant face representations.It explicitly reduces the difference between source and target domain by fine-tuning the model with extra MK-MMD loss and adversarial network,which makes the model more robust to open scenes.Finally,this paper designs scientific and reasonable experiments to verify the performance of each module.In the experimental stage,the effectiveness of the proposed methods are proved by the comparison with state-of-the-art algorithms and self-control analysis.Our proposed algorithms achieve great performance on both public and private testing data sets,which verifies the effectiveness.