Research On Heterogeneous Face Synthesis Based On Generative Adversarial Networks

In the field of digital entertainment and public security,the synthesis of heterogeneous faces plays an increasingly important role.Heterogeneous face refers to the face image with different forms,including sketch face portrait,cartoon face image,visible face image,near infrared light face image,etc.Heterogeneous face synthesis aims at transforming face images into different morphologies to meet the actual needs.However,most of the previous methods of heterogeneous face synthesis are carried out in a supervised way,and the quality of the synthesized image is not high.Therefore,aiming at the problems in the synthesis task,this thesis takes the most common face image synthesis task of sketch and visible light photo as an example,and constructs a synthesis model of sketch realistic to photo and a synthesis model of visible light photo sketch stylization based on GAN.The specific work of this thesis is as follows:First,in order to solve the problems of missing facial details,insufficient realism,and artifacts in the synthesized face images of the current face sketch to photo method,a model based on unsupervised synthesis of face sketch images to photos is proposed.Firstly,the multi-scale feature extraction module is used to extract richer semantic information from the source image.Secondly,the convolutional attention module is embedded in the generator to enhance the ability of the model to extract important feature information,improve the quality of the synthesized image and reduce the artifacts caused by the background interference of the image.In addition,in order to make the generated face image retain more key details,facial detail feature loss is designed to be applied to the model to constrain the generation of local details.The results of qualitative experiments on CUFS and CUFSF public data sets show that the facial details and edge structures synthesized by this model are clearer and more realistic.At the same time,the performance indexes such as structural similarity and peak signal-to-noise ratio are significantly improved compared with other methods.Second,in the field of digital entertainment,photo sketching also has great application prospects.In view of the problems such as rough texture and not fine enough in the traditional face sketch synthesis method,the residual dense network is used as the main structure of the network to make full use of the extracted feature information.Secondly,the more robust Charbonnier function is used as the pixel cycle consistency loss function to optimize the processing of outliers deep in the network and reduce the training time.The citation Tv loss further removes the rough texture present at the edges of the composite sketch image.At the same time,the structure of the discriminator in the training process is optimized,and the field of view of the discriminator is enlarged and the discriminant ability is improved.The qualitative and quantitative experimental results on CUFS and CUFSF public data sets show that this method can obtain more real and delicate sketch face composite images.