The Research Of Face Frontalization Based On Generative Models

With the widespread construction and application of video surveillance network in China,the rapid identification of pedestrians through video surveillance has become a key factor in the intelligentization of public security work.In reality,most of the faces obtained by video surveillance are non-frontal faces,which affects the accuracy of face recognition and subsequent use.Face frontalization provides an effective preprocessing method for multi-view face recognition.Face frontalization refers to generating frontal faces based on profile faces,so can significantly improve the multi-view face recognition accuracy.The synthesized frontal face can also be used in the subsequent public security work.So face frontalization has a broad application prospect.This paper proposed new methods of face frontalization based on the generative model.Through the research and analysis of the current generative models using deep learning technology,both generator and discriminator of generative adversarial learning are improved to better handle the problem of face frontalization.We design two generative models that can better achieve the task of face frontalization.The main work of this paper is as follows:Firstly,we improved the generator network by respectively setting VAE model and GLOW model as generators.For the generator network based on VAE,a conditional ?-VAE model is proposed to enhance the disentanglement ability of face generation.For the generator network based on GLOW,the proposed conditional flow model takes advantage of the recently introduced conditional batch normalization CBN layer to control the direction of generation.Secondly,we improved the discriminator network by respectively adding the identity,pose loss function and mutual information restriction to improve the objective function.The discriminator network with identity and pose loss function was no longer a traditional binary classifier,but a multi-output classifier.It can not only predict the type of true or false,but also predict the identity and pose of face pictures.For the discriminator network with mutual information restriction,the mutual information of the original data and the latent variables was restricted to balance the training of the network.Thirdly,we proposed two face frontailzation models.They were the ?VAE-GAN model with conditional ?-VAE generator network and the added identity and pose loss function discriminator network and the Cflow-GAN model with conditional flow generator network and the added mutual information restriction discriminator network.By taking advantage of the generative adversarial framework,we realized the task of face frontalization.Fourthly,We analyzed the proposed two face frontalization models through various evaluation methods.In the quantitative analysis,?VAE-GAN and Cflow-GAN were better than the benchmark model under evaluation standards such as FID,IS,PSNR and SSIM.In the experimental evaluation of face frontalization recognition,the recognition accuracy of ?VAEGAN and Cflow-GAN was improved.Based on the above work,this paper proposed two face frontalization models and analyzed them through various evaluation standards.We proved that the two proposed models can generate high-quality frontal faces,which improves the accuracy of multi-view face recognition.