Research On Fine Controllable Face Editing

In the field of computer vision,the study of image generation model is an important and interesting direction,based on the search and calculation of the real image information distribution,the realistic results satisfying the target feature distribution are generated.Thanks to the powerful quasi-cooperation of deep learning,it makes it possible to learn complex features.Generative Adversarial Network(GAN)is one of the outstanding models of image generation.Most of the research work is based on GAN.This paper takes face attributes as the condition to generate realistic images that meet the requirements,deeply studies the algorithm of attribute editing and attribute interception,and improves the quality and diversity of generated images.The research contents of this paper are as follows:(1)For the current unsupervised attribute editing network,it can be divided into tag-driven generation and reference graph-driven generation.This paper makes an indepth study on whether the dual-driver fusion into one model can effectively improve the model performance and its reasons.For the unique formal structure of dual drive mode,this paper designs a new network structure and loss function to narrow the distribution of the two drives in the hidden layer space,give play to the advantages of dual drive,further effectively edit attributes,and improve network performance.(2)In this paper,we propose a new task Angle for age and perspective attributes in face attribute editing task,and divide them into a continuous domain of coherent feature fitting.For the age editing task,this paper considers various possibilities of the actual aging process,and from a new research perspective,abandonsthe traditional approach of over-pursuing age accuracy,and puts forward the age diversity of the generated images for the reality and users,giving users more choice space.For most age editing tasks,extreme ages cannot be further edited to both ends in the reasoning stage.This paper proposes a conversion module to solve the functional defects of this model.