A Research Of Emotion Synthesis Based On Generative Adversarial Networks

The continuous research of facial emotion synthesis technology has made a longterm progress in graphics,video and animation.Human facial features provide rich information,which is an important part of affective computing and identity recognition.In real life,facial emotion synthesis technology has a wide range of applications,such as entertainment film and television production,computer animation synthesis,and medical beauty simulation.Thanks to the development of image deformation technology and the development of depth neural network model,the complex operation of artificial modeling can be replaced by building generative model,so facial emotion synthesis technology has a broad development prospect.The main work of this paper focuses on the difficulties in the research of emotion synthesis technology.This paper improves three aspects of emotion synthesis.One is the continuous emotion synthesis control method;the other is the method supporting the transformation between categories(different emotions)and different types(different styles);the third is the quantification method of image effect evaluation and the optimization strategy of synthesis quality.For the traditional facial animation synthesis technology,the cost of money and manpower is high,and later modification and maintenance is difficult.In the framework,we adopt the structure similar to the cyclegan,a kind of generation adversarial network,and try to add the control of emotion in different parts of the structure.The structure designed in the model parameters is simple and runs fast.The generation adversarial network creatively combines the generation model and the discrimination model to synthesis data,but the generation mode of GAN is too free,and it is easy to lose control when there are many image pixels.To solve this problem,a weak supervision control mechanism is designed,and this control is a continuous control.In the process of the experiment,because of the privacy and spontaneity of the emotion database,the amount of data is small,and most of them are recorded in the laboratory environment,and the differences between different databases are relatively large.We use OPENFACE and face+ + tools to create a large number of databases for training emotion generation,which is convenient for the follow-up research.At the same time,we give the quantitative method of image effect evaluation,both subjective and objective comprehensive evaluation way,for the reference of follow-up personnel.Finally,this paper compares the differences of several corresponding emotion generation frameworks.The framework of this paper has advantages in continuous control and cross domain transformation.Experiments are carried out on ffhq,OuluCASIA and danbooru2018 databases to verify the effectiveness of the framework on subjective and objective indicators.