Research On Face Attribute Editing Algorithm Based On Generative Adversarial Network

With the development of the application of computer vision technology,face attribute editing has been continuously expanded in many applications.It not only strengthens the existing basic functions such as face microdermabrasion and face whitening,but also has been widely used in the fields of live video and face confrontation attacks.Face attribute editing aims to change one or more attributes of the image,such as hair color,skin color,age,etc.,and keep other attributes unchanged.While editing the attributes,it is necessary to maintain the high quality and accuracy of the target attribute images.The development of generative adversarial networks in recent years has greatly promoted the evolution of face attribute editing algorithms,but the existing algorithms still have some shortcomings.Most algorithms focus on editing with or without certain attributes,and lack the addition and deletion of attributes for specific templates,such as the addition of bangs with specific shapes.On the other hand,when face multi-attributes are edited at the same time,it is difficult for existing algorithms to remove the interference between multiple attributes in the feature space.This will result in more artificial artifact effects and background changes in multi-attribute editing.In response to these problems,this paper studies the face attribute editing algorithm based on the structure of the generated adversarial network.The main work of this article is as follows:(1)A face template attribute editing algorithm based on two-way coding and attention mechanism is proposed.The existing template attribute editing algorithm,such as the ELEGANT model,lacks consistency in the attribute style of the generated graph and the reference template graph,and is prone to defects such as artifacts.In response to this problem,this paper uses an independent attribute encoder and background encoder to encode the image in two ways,disassociate the attribute correlation vector and attribute independent vector,and improve the independence of attribute representation.On the other hand,in the decoding process,the attention multidimensional feature fusion method is used to filter and combine the background multi-dimensional vectors with a lot of details and introduce them into the decoding operation,which is conducive to improving the quality of image generation and the accuracy of attribute editing.Finally,the experimental results on the Celeb A data set show that the algorithm can achieve a higher image quality and more accurate attribute template editing face generation results.(2)A multi-attribute editing algorithm based on recurrent network is proposed to solve the problems of inter-attribute interference and low quality of the existing algorithms in multiattribute editing.On the one hand,after extracting attribute features and background features with an encoder,for the purpose of dissociating multi-attribute features in the feature space,this chapter introduces attribute condition tags to participate in hidden layer exchanges,and proposes an operation strategy for hidden layer multi-attribute de-correlation,which is beneficial to upgrade The representation independence of each attribute sub-vector in hidden layer space.On the other hand,in order to improve the naturalness of the multi-attribute editing results and reduce the artifacts and the destruction of the face structure,this chapter uses a recurrent network architecture to introduce a feedback loop on the existing basis for extracting hidden layer output and output images Feedback characteristics.The experimental results show that the algorithm can generate clearer and higher-quality face editing images,which has a significant effect in the face multi-attribute editing task.The research content of this article can take into account the specific template attribute editing and multi-attribute simultaneous editing,which is practical in real scenes.