Face Age Synthesis For Weakly Labeled Data

Facial age synthesis is a technology to study someone's past and future appearance.This technology has a wide range of applications in real life,such as searching for missing people,film and television entertainment,public welfare project publicity and so on.But as we all know,face images contain a lot of important information,such as gender,expression,identity,age and so on.How to obtain these information and how to retain more effective personal information are worthy of further study.In addition,due to the scarcity of data,the research in this field is facing great challenge,and how to solve the problem of insufficient data is also a direction for further study.In view of the above problems,this paper makes contributions as follows:(1)This paper proposes a face age synthesis method,which encodes the face image through the conditional variational auto-encoder technology,combines the face feature vector with the target age,and then synthesizes the face image according to the age condition through the generation network.This method has been trained and tested on existing datasets,and can generate good regression and aging face images.(2)This paper proposes a training method based on weakly labeled data,which uses semisupervised learning technology to learn face features and complete the training of face age synthesis model based on weakly labeled data.Compared with the current face age synthesis method,this method can solve the problem of inaccurate label and lack of pair-wised face image data while completing the task of face age synthesis.(3)This paper proposes a face age synthesis model based on cycle optimization algorithm.In this model,the face image is first obtained by the encoder,and then the new face image is generated by adding the age condition,and then the new face feature vector is obtained by inputting the encoder.By introducing the loss function of personal information,more identity information can be retained.At the same time,gender condition is added to the model for gender grouping training,so that the generated regression and aging face images can be improved.Extensive experimental results fully verified the advantages of this model compared with other existing models.The model in this paper can generate regression and aging face images with more realistic vision and better effect,which can be applied to different scenes.