Facial Expression Transformation Based On Generative Adversarial Networks

Facial actions are the main way for humans to convey information on a daily basis,and they bear the important responsibility of emotional transmission.Therefore,the recognition,analysis and synthesis of facial actions are essential technologies for achieving human-computer interaction.With the continuous development of the field of deep learning,the processing technology of facial images has gradually matured.Image synthesis based solely on character identity can no longer meet the application requirements,generating more realistic facial motion,and simulating real facial expression has become one of the important resear-ch directions in the computer field.With the development of these key technologies,facial expression transformation combined with various technologies has emerged.Facial expression transformation technology has been applied in many aspects such as privacy protection,biomimetic agent,special effects synthesis,data expansion,etc.,and has important application value.Based on the analysis and summary of the existing research,we found that there are two problems in the current research on expression migration.First,there is no optimal method for the representation of expressions,where descriptive representations carry subjective information and action coding carries personal habit characteristics.Both of them have had a certain impact on the later processing.Secondly,in the process of migration,there is a lack of treatment of personal habit characteristics.The defect that caused the expression to not satisfy the actual character situation.To solve these problems,the main research contents of this paper are as follows:Firstly,this paper first proposed a general expression representation method,which is called an expression base.Through the mapping relationship between the facial action unit and the two-dimensional discrete expression,the universal features jointly expressed by the two are calculated to describe the expression,thereby reducing the subjectivity of the discrete expression and the specificity of the facial action unit.First,through a two-dimensional discrete expression tag recognition model,all the data in the training set is reclassified by the model to unify the subjective error of the two-dimensional discrete tags.Then,the common action unit sequence corresponding to the two-dimensional discrete expression tags is calculated according to the reclassified data,the personal habit characteristics are eliminated,and an expression base that can be used to describe the universal expression action is obtained.Secondly,aiming at the process of expression transformation,this paper proposes an expression-based based feature conversion algorithm.Using the expression base as the conversion medium,the habit factor matrix and the habit difference matrix are designed as conversion parameters,which represent the difference between the expression change amplitude and the expression change of the target personality.First,using the expression base as a benchmark,learn the habit factors and habit differences of the target person from the target person image.Then,at the time of conversion,the two-dimensional discrete expression of the source face image is recognized to obtain its corresponding expression base.Finally,through the two matrices,the sequence of corresponding facial action units containing the personal habits of the target person is calculated,and the transformation of the habitual features is realized.Thirdly,a personalized facial expression transformation system based on generating a confrontation network was designed and implemented.The conversion calculation for personal habit characteristics is realized,which solves the problem of neglecting habit characteristics in facial expression transformation research.The validity and rationality of the system are verified by subjective and obj ective evaluation methods.