Research On Face Image Synthesis Based On Generative Adversarial Networks

Face image synthesis has always been a popular and difficult research issue in computer vision and graphics.Face image synthesis is to generate a high realistic face image according to the input information,which involves many research issues,such as face expression transfer,face attribute editing,and face image restoration.In recent years,face image synthesis researches,which are based on deep learning and generative adversarial network in special,has drawn the attention of many scholars.The technology is widely used in the fields of animation production,digital entertainment,security prevention and control,etc.Due to the particularity of human face,subtle changes in the face image will bring great visual differences,and the transmitted emotions are also very different.Therefore,the face image synthesis technology mainly faces the following challenges: first,how to improve the controllability and diversity of face image synthesis,so as to obtain a face with multiple appearances and rich expressions that meet the expectations of users;the second is how to make the synthetic face keep the given identity well,and make the expression more real and natural;the third is how to optimize the synthesis efficiency and generalization ability of the face synthesis model.For example,once the model is trained,the conversion between multiple characters or multiple expressions can be realized without retraining the model;fourth,how to make full use of the high-dimensional and low-dimensional information of the given input face image to make the facial geometry of the generated face more realistic and reasonable,and the texture information richer.To tackle the above problems,based on the generative adversarial network,this paper studies the high realistic face image synthesis technology from two aspects: face synthesis combined with expression transfer and appearance editing and face image reconstruction.The contributions of our research are as follows:1.We propose a novel face synthesis network model driven by facial geometric feature and attribute label.Given a source face image,target face image and the attribute label,the new face synthesis model can generate a highly realistic face image which owns the expression of the source face,the identity of the target face and the specified attribute.A novel soft margin triplet perceptual loss function is introduced to the model,which can make the synthesized face more real and natural and keep the identity of the target face well.The proposed model has the functions of expression migration and appearance editing,and allows users to meet different application needs by controlling input information.For example,without the input of the attribute label,the face synthesis model can transfer the expression from the source face to the target face,and the generative expression is true and natural,and keep the identity character of the target face;with the attribute label input,the face appearance can be edited according to the specified attributes by the model,where the expression is transferred,and the high realistic face image with diversified appearance and expression can be synthesized.2.We propose a face image reconstruction model based on facial geometric and texture prior to solve the problems of unclear facial features and loss of facial texture details and improve the quality of generated face image.Given a low-quality face image,the new model can generate a high-quality face image with clear facial features and hairstyle,rich texture details and consistent with the identity characteristics of the input image.The model introduces the idea of geometric prior,the face parsing map prediction module can better retain the geometric information of the low-quality input face,so it can retain more geometric features of the input face.The GAN prior decoder with rich facial texture information can reconstruct the facial details of the input face.The combination of geometric prior and texture prior retains facial geometric and texture details.In addition,the multiscale encoder in the model can integrate the high-dimensional and low-dimensional information contained in the input face at different scales,and retain the characteristics of the input face image to improve the quality of the reconstructed face.We propose a face synthesis network model driven by geometric feature and attribute label and a multi-scale face image reconstruction model based on facial geometric and texture prior,which improve the generalization ability of the model and make the generated face image more realistic.