Research On Facial Age Synthesis Method Based On Style Fusion And Selective Domain Discrimination

Facial age synthesis aims to use computer technology to analyze the features of facial images,then synthesize face images that meet the target age.It can be applied to film and television production,face recognition,portrait prediction,forensics,etc.With the development of image processing technology,the facial age synthesis method has gradually changed from traditional computer graphics method to deep learning method based on generative adversarial networks,and the quality of generated image has taken a qualitative leap.However,most existing facial age synthesis methods fail to take personalized face attributes and features into consideration,and the number of samples in different age groups in such datasets are unevenly distributed,which makes the training process unstable and seriously affects the actual effect of age synthesis.To address the above problems,this thesis proposes a facial age synthesis method based on style fusion and multi-domain discrimination.The main research contents of this thesis are as follows:The proposed generative adversarial network based on style fusion and selective domain discrimination includes a generator model based on style fusion and a multi-domain discriminator with selective age domains.The generator is responsible for extracting the identity features of the source image and generating a face image that meets the target age condition.The discriminator is responsible for stably providing the generator with feedback information for different age domains,and guiding the generator to learn the mapping relationship with different age domains.In the generator network based on style fusion,the condition mapper module maps the target condition label to the condition style,and the identity extractor module extracts the identity style of the source image.The condition style and the identity style are fused to modulate the decoder features,in order to make the generated target face preserve the personalized identity features of the source image.In the multi-domain discriminator based on selective domain structure,the feature extraction modules with different scales adopt the selective domain structure to extract general features of different age domains and discriminative features in specific target domain individually,then the two kind of features are combined to make up for the lack of feature distribution in the age domain with fewer samples,thus to alleviate the problem caused by the unbalanced training data,and then fit the feature distributions of different age domains better.Extensive subjective and objective experiments have been conducted on the FFHQAging,Morph and CACD datasets,and the experimental results show that the proposed facial age synthesis method can preserve the personalized identity information of the source image,generate facial age features that meet the target conditions,and robustly simulate the feature distribution of different age domains.