Research On Face Generation Algorithm Based On Generative Adversarial Networks

Facial image generation is one of the basic tasks in the field of computer vision.It realizes the cross domain and diversity generation of images.With effective combination of Generative Adversarial Networks,style transfer technology and image editing technology,the more complex facial generation tasks such as facial style,facial makeup transfer,human pose transfer and facial expression are realized.Now,the diversified control network is added to realize clearer face image generation and more intelligent style effect in Generative Adversarial Networks.By systematically analyzing the overall training process of facial generation algorithm based on Generative Adversarial Networks,and considering the network structure and training strategy of Generative Adversarial Networks,this paper proposes more efficient facial generation algorithm.The main work contents are as follows:Aiming at the problem of low-level attribute generation in face generation algorithm based on Generative Adversarial Networks,this paper,analyzes the causes of the problems,and puts forward asymmetric input face generator,style code and scalable noise to effectively control face detail features,adjust low-level attributes,maintain the generated image and improve the generation quality.With the improvement of generation ability,the feature quantization discriminator of dictionary structure is used to improve the discrimination ability of discriminator.The two network structures interact with each other to jointly improve the performance of the generation model.At the same time,the dictionary loss function and the identity loss function stabilize the whole algorithm training process.Experiments will prove the effectiveness of the algorithm and the improvement of generation ability.Aiming at the low quality of face generation in the face generation algorithm based on Generative Adversarial Networks,the training strategy of contrastive learning mode and the adversarial training strategy of reconstructed image are proposed to improve the quality of face generation.The training strategy of contrastive learning mode includes contrastive learning discriminator and conditional contrast loss function.At the same time,in order to avoid limiting the advantages of the training strategy of contrastive learning mode to the discriminator,the adversarial consistency loss function is proposed,which can not only let the discriminator participate in the image reconstruction process of the generator,and the cross entropy calculation method can further ensure the integrity of the face and make the model more robust.