Face Image Synthesis And Editing Based On Generative Adversarial Networks

With the rapid development of computer technology and the emergence of big data,artificial intelligence technologies such as machine learning and deep learning,Image synthesis technology and image translation technology based on deep learning have achieved milestone progress.For example,Style GAN based on Generative Adversarial Networks(GAN)can synthesize high-resolution face images that are difficult to distinguish with the eyes.The day a software named Deep fake has been proposed,it has aroused academic and industry scientist's widespread concern.Among many face image synthesis and attribute editing models based on deep learning,GAN has been widely concerned because of its outstanding performance and the idea of adversarial learning.However,the current modes still face many challenges and problems in theory and application.For instance,Wasserstein GAN(WGAN)points out that when discriminator converges rapidly,the gradient of generator will disappear rapidly,which is the essential reason of the instability of GAN training.In addition,in the multidomain face attribute editing model,there are also problems such as attribute entanglement.To solve the above problems,this paper focuses on the two tasks that face image synthesis and face attribute editing based on GAN.The main contributions are as follows:We proposed Weighted Residual Network based on generative adversarial network(WR-GAN)for face image synthesis.Inspired by the deep convolution GAN(DCGAN)to improve the quality of image synthesis by modifying the network architecture,this paper proposes a weighted residual network for image synthesis.The experiment results show that the weighted residual network can improve the quality of the image synthesis.In addition,as shown in the WGAN,the discriminator of GAN converges rapidly,which will lead to the instability of GAN training.In order to solve this problem,cyclic learning rate is used to repleace the original learning rate of discriminator .The experiment results show that it can not only effectively solves the problem of fast convergence of discriminator but also helps to improve the quality of image synthesis.We propose a face attribute editing model based on multi-path dual attention mechanism generative adversarial network.For multi-target domain image translation tasks,especially face attribute editing,the output features of the encoder should not only be able to realize the conversion between different domains according to different input label,but also ensure that all the attributes except for the label remain unchanged.Therefore,when the amount of data is small,the phenomenon of attribute entanglement will appear,as shown in section 3.4.5.To solve this problem,we emply a multi-path dual attention mechanism.The experiment results show that it is not only conducive to alleviate the problem of attribute entanglement,but also improve the quality of face attribute editing.In addition,in order to eliminate the difference in semantic features between the generated image and the real image,this paper introduces the perceptual loss in the face attribute editing,and the experiments show that the perceptual loss is helpful to the face attribute editing task.