Face Pose Reconstruction And Faec Age Editing Research Based On Self Attention And Multie Feature Discrimination

The face is one of the important biometric features,and it is important to generate high-quality pictures of edited properties.In recent years,Many significant research results have been produced in the field of face frontalization and face age editing.Nevertheless,the frontal reconstruction of large-scale profile faces still confronts the difficulty of incomplete facial information,and in the field of face age editing,the lack of image data of the same person of different ages will lead to the fact that the facial details generated by models are not realistic.Therefore,the research on face-based pose reconstruction and age attribute editing is still challenging.The existing mainstream methods of face frontalization mainly focus on optimizing the generator structure and loss function,however,improving the performance of the identification network is ignored.The factors mentioned above result in single extracted features and blurred image details.The research on face aging focuses on deepening the layers of the network model and enriching the input of guidance information but does not consider the strengthening constraints on the face texture features,which result in the lack of facial details for the generated target age face,and the image quality is difficult to further improve.Aiming at the problems in the above research,this paper proposes a face frontalization network based on variable-scale self-attention and a face aging network based on multi-feature discrimination.The main research work is as follows:(1)Evaluating the existing mainstream methods of face pose reconstruction,face aging,and attention mechanism.Analyzing the technical route and innovative ideas of each method.Finally,summarize the shortcomings of each method and proposed feasible improvement plans.(2)A frontal face reconstruction method based on detail identification,variable scale self attention and flexible skip connection(FR-DVF)is proposed: Extracting the multi-scale local information of the generated faces and sending them to four independent discriminators for authenticity identification to improve the image quality;Adding the variable-scale self-attention mechanism in the generator to strengthen the extraction capacity of spatial context features and reduce the dimension of the spatial correlation matrix dimension;Introducing flexible skip connection to enhance the compatibility of cross-layers to improve the compatibility between the feature layers.(3)A face age editing method based on pixel self attention,multi-feature discriminated and pretrained-feature-fusion,AE-PMP): Introducing a pixel attention module with direct allocation of full pixel weights,which strengthens the network’s ability to extract context space features;using filtering operators to extract edge gradient information of pictures and send them to multi-feature discriminator groups for judgment to enhance the network’s ability to perceive edge gradient features;In the generator,the pre-trained model and the encoder simultaneously extract facial features and perform channel splicing in the hidden layer,which improves the texture details and diversity of the generated images.Experimental results show that the face frontalization network proposed in this paper can reconstruct frontal faces with clear details.On the M2 FPA dataset,when the Rank-1 recognition rates(%)are used as the evaluation standard,FRDVF is 1.53% higher than CAPG-GAN.When use SSIM and PSNR to evaluate the pictures’ quality,FR-DVF outperforms CAPG-GAN by 1.27% and 0.86%,respectively.The proposed AE-PMP face age editing algorithm can generate a realistic look while maintaining the original identity features.Taking the predicted age indicator as an example,the age prediction accuracy on the FFHQ and Celeb A-HQ datasets is 5.20% and 3.21% higher than that of IPCGAN,respectively.To sum up,for the problem of face pose reconstruction and face aging,this paper proposes FR-DVF and AE-PMP networks respectively.The experiments on public datasets confirm the effectiveness and superiority of the two networks.Finally,this paper discusses the shortcomings of the two models and proposes and propose feasible improvements in the future.