Research And Implementation Of Face Generation Method Supporting Pose Customization

Computer Vision is a discipline that focuses on simulating human visual perception,analyzing and deeply understanding image and video data.Image Generation is a new research task in computer vision,which obtains generated images with higher fidelity by learning the distribution of real pictures and sampling from the learned distribution.This paper focuses on the face generation in the field of image generation,and studies the face generation methods that can support pose customization.Thanks to the current generation model's fitting and modeling of face distribution becoming clearer and more realistic,this paper studies a variety of innovative variant algorithms based on the high-definition face generation model to make it more controllable and explainability,ultimately it has a wider range of application scenarios and more application value.Several important functional sub-modules explored in this article are:face projection,face coding,face editing and expression migration.The related research results can be applied to applications such as making short-videos.Based on StyleGAN,this paper proposes a variety of algorithm frameworks and adaptation methods through appropriate improvements and expansions,corresponding to the four functions of face generation,face embedding,face editing,and expression migration.The main work of this paper includes:(1)Proposing the Info-StyleGAN framework to realize face generation of control categories,(2)Applying the transfer learning method to StyleGAN to show its rapid adaptation effect,(3)Proposing MLP-ResNet50 to achieve a better face embedding effect,(4)Applying the fast latent space interpolation method to StyleGAN to show high-quality face editing effects,(5)Proposing that FET-StyleGAN can be used for expression migration tasks.In addition,this article also shows a set of web version face editing system that supports posture customization.The value of this article is to explore the editing system of high-definition face generation(1024x1024)and propose a variety of operation algorithms with good controlbility.Compared with the conventional generation model(256x256 resolution),this research is more difficult but has more application value.