Research On Facial Attribute Transfer Based On Attribute-Independent Generative Adversarial Networks

Face attribute transfer is an image synthesis technology.It refers to generating a face image with the same identity information and specified attribute information from a real face image while preserving other details.The face image generated by the traditional method based on feature extraction neural network has poor realism and lacks detailed information.In order to improve the quality of generated images,generative adversarial network is introduced into the face attribute transfer algorithm.The main problems in the existing face attribute transfer algorithms include:some algorithms consider the independence among attributes but ignore the integrity of the information,leading to the loss of identity information in the generated image;others consider the integrity of the information but the independence among attributes can not be guaranteed,leading to changes in non-specified attribute information in the generated image;the discriminator and generator in GANs are difficult to reach the convergence state at the same time,leading to unstable training,mode collapse and so on.Based on the above-mentioned problems of the existing methods,this paper proposes a facial attribute transfer algorithm based on attribute-independent generative adversarial networks.At the same time,a variable alternating training mechanism is proposed for discriminators and generators in GANs.The work of the thesis is divided into three parts:(1)A facial attribute transfer algorithm based on attribute-independent generative adversarial networks is proposed.This thesis introduces a hierarchical structure of identity vectors.Different parts of identity vectors can control the different levels of features in the generated face image.The structure of the codec is introduced into the network,and the reconstruction loss of the identity vector is introduced into the loss function.This algorithm can not only guarantee the independence among attributes,but also minimize the loss of other detailed information in the generated face images.(2)A training method for generative adversarial networks is proposed.In addition to solving problems such as unstable training and mode collapse,this method also proposes a variable alternating training mechanism for discriminator and generator.During the training process,the values of the alternating parameters can be adjusted in real time according to the current state to maintain dynamic balance of discriminator and generator.This ensures that the network can obtain a higher gradient value in the backpropagation to effectively prevent the gradient vanishing.(3)The effectiveness of the above work is verified by experiments,which mainly includes a comparative analysis with existing facial attribute transfer models in generating face images and a comparative analysis of the convergence of GANs with different training methods.Experiments show that the facial attribute transfer algorithm proposed in the paper can generate more realistic and clearer face images;the training method proposed in the paper can accelerate the convergence process of the network while preventing the gradient vanishing.