Image-to-image Translation Based On Structure Improved Generative Adversarial Networks

Image translation algorithm refers to the process of giving an image and generating another image through certain mathematical methods or computer processing,which widely applied in fonts transfer,attributes transfer and saliency prediction.In this paper,we focus on facial local attributes transfer and web page saliency prediction task in image translation.Facial local attributes transfer is an image translation task that pays more attention to local information.The part area of image needs to be translated while identity information needs to be kept unchanged.Web page saliency prediction is an image translation task that pays special attention to web page edge information.Since web pages usually have high resolution and each area in web page image is independent with each other,it is difficult to realize image translation by one step prediction.The generative adversarial nets(GAN)is one of the most important research field of deep learning that can effectively realize image translation.In view of the characteristics of facial local attributes transfer and web page saliency prediction,we designed two sets of structure improved generative adversarial networks and verified the validity of the models.The main work of this paper is:Firstly,considering the special importance of local information in the facial attributes transfer task,this paper improves the discriminator part of the generative adversarial networks and proposes Dual Discriminative Adversarial Networks(DDAN).The model introduces another local attributes discriminator based on the original generative adversarial networks to form dual discriminative structure.The two sets of discriminators extract the global and local information of the image respectively.The model also introduces an identity information retention network to ensure that the identity information of the images before and after the attributes transfer is unchanged.We validate the effectiveness of DDAN on the five local attributes transfer tasks on the CelebA datasets.The image classification accuracy rate transferred by DDAN is significantly better than baseline models.Secondly,considering the difficulty of one-step prediction in web page saliency prediction,this paper improves the generator part of the generative adversarial networks and proposes Dual Generative Adversarial Networks(DGAN).The model realizes a coarse to fine web page saliency prediction by introducing a coarse generator and a fine generator,and changes the traditional one-step web page saliency prediction into twostep prediction process.The model also introduces the edge information of the web page extracted by the Laplacian operator,which improves the predictive effect of the model.We verified the effect of the DGAN in the web page saliency prediction on the web page datasets FIWI.CC and NSS score shows that DGAN is significantly better than baseline models.