Research On Temporal Face Synthesis Model Based On Transfer Adversarial

Face features have become an increasingly important feature information in daily life,and the feature of faces will slowly change over time.Human facial aging is an irreversible process,specifically manifested in changes in facial texture,shape,hair color,and beard.How to generate a face aging image that conforms to the laws of nature has become a hot issue in recent years,and it has also become one of the key research directions of many fields.In the field of temporal face synthesis,models based on physical methods and prototype methods have appeared successively.After the generative adversarial network was proposed,more and more methods based on the generative adversarial network were proposed.However,due to the limitations of cross-age face data sets and model structures,the quality of face images generated by existing methods is not high and there is a lack of research on Asian face data sets.In view of the above problems,for the single-task time-series face synthesis scene of synthesizing face images of a specific age domain,this thesis proposes a single-task timeseries face synthesis model based on transfer adversarial.This thesis uses transfer learning to redesign the structure of the generation network and the structure of the discrimination network in the GAN and improve the loss function,so that the model can better learn the feature information in a small amount of data set in a single-task scenario.In another common multi-task scenario that requires the synthesis of multiple face images of different age domains,this thesis proposes a solution to the problems of existing models,such as the lack of research on Asian face data and the low quality of the generated images.A multi-task temporal face synthesis model based on transfer adversarial.This thesis introduces an auxiliary discriminant network structure based on transfer learning.Through the auxiliary discriminant network,generated face image can be mapped into the face feature space,and the age domain label of the face can be generated.Through intuitive comparison of facial features and age domain labels,the quality of the images generated by the model is improved.Experiments and analyses are carried out on the Asian face dataset.At the same time,in another different age domain scene,experiments and analysis were carried out on the CACD data set.Extensive experimental results have demonstrated that,compared with the existing models,the proposed model performs better under objective evaluation indicators.And the scalability and diversity of the model are better...