Research On Face Aging Simulation Method Based On Generative Adversarial Network

Face aging is an important research content in computer vision.By age mapping,the face with original age is transformed into the face with target age,which can provide support for face recognition,face animation and other applications.Age mapping can lead to changes of facial texture and facial shape.During the period of age mapping,one problem cannot be ignored.It is that no matter whether the old face is transformed into the young face or the young face is transformed into the old face,the inherent identity information of the face must be keep.That is,the age mapping cannot lead to the transformation from one person to another.In this regard,the current main method is to decouple the identity information and the original age information,and then inject the target age information.However,if the age span increases,the decoupling will lead to the weakening of the identity characteristics that represent people.The larger the span,the more obvious this problem is.In addition,age mapping will also cause changes in texture and shape,affecting the effect of the target age face.The subject work is carried out around these two aspects.To solve the problem of completing age mapping and ensuring identity consistency,a face aging method based on age difference is proposed.This method does not decouple identity information and original age information,and adds target age information on the basis of removing original age information,but adds age difference information on the basis of original age information.Through the learning process,the mapping from original age face to target age face is completed.In order to avoid the destruction of identity consistency,identity consistency constraints are imposed during learning.The learning process uses a generative adversary network,which uses age coding to describe the age.The generator uses the encoder and decoder structure.The encoder completes the feature extraction of the face at the original age.At the decoder stage,the original age coding is not removed directly and the target age coding is added to avoid decoupling identity information and age information.A new age code is designed,age segments are divided,and every 50 bits are used to represent the age difference information between two adjacent age segments,so that the age change becomes a continuous process,and the increment of each age change is controlled within a certain age range,so that identity consistency can be further maintained.The age difference truth value representation is constructed,and the decoder process is introduced to make the age difference coding contain the real age difference information between different age groups.The discriminator part of the generative countermeasure network adds identity consistency constraints.The experimental results show that this method can achieve better aging effect,and the generated face is closer to the actual appearance of the person at the target age in the image.The problem of texture and shape changes caused by age mapping is studied to achieve fine control of age changes.The facial key points are optimized by joint alignment,and the facial key points in the shape standard face and input face are obtained.The Euclidean Distance is calculated for the face key points to obtain the shape standard face key points of the original age group and the target age group,which are closest to the input face key points.The shape difference model is constructed to describe the shape change of the target age.The Euclidean Distance between the texture standard face of different ages and the input face age is calculated to obtain the texture standard face closest to the input face in the original age and target age.Texture difference model is constructed to describe the texture changes of the target age.The texture difference and shape difference of the target age and the original age are inject into the age mapping to synthesize an image close to the actual face.The experiment verifies the performance of joint alignment method,and the aging model after introducing shape difference and texture difference.