Research On Generative Adversarial Network And Its Application In Graphic Design

The Generative Adversarial Network(GAN)was first formally proposed by the research team led by Goodfellow in 2014.It contains a pair of competitive models of generator and discriminator.By adopting a unique confrontation training mechanism for the two models,the generator can learn the latent distribution of real samples without priori assumptions.Since the birth of GAN,GAN has been highly concerned by academia and industry.On the one hand,there are many improved models that have been proposed one after another for GAN ’s difficulties in training and mode collapse.On the other hand,given that GAN’s framework is very flexible,there are many studies applying it to different fields.Based on these two aspects,we conduct in-depth theoretical and experimental research on GAN and its series of improved models.In addition,we expand the application of GA N to graphic design.The detailed research work done in this thesis is summarized as follows:Firstly,through comprehensively combing the relevant research of GAN,five typical models for improving the original network structure or loss function are selected,namely DCGAN,WGAN,WGAN-GP,LSGAN and CGAN.We focus on the research contents of GAN and these five improved models,and first analyze them in detail in theory.Next,in order to compare generati ve capabilities of the six models,we collect 104383 anime character images from the Internet,combining with the characteristics of the dataset and the network structure of DCGAN,we have made appropriate improvements to the network structure of WGAN,WGAN-GP,LSGAN and CGAN,then use these six models to cond uct image synthesis experiments on this dataset.The experimental results show that the effects of the images generated by WGAN,WGAN-GP,LSGAN,and CGAN are similar,and they are generally better than GAN and DCGAN.In addition,the effect of the images generated by DCGAN is better than GAN.The experimental results verify the superiority of the loss functions of WGAN,WGAN-GP,LSGAN and the effectiveness of the network structures of DCGAN,CGAN.In view of the serious training problems in generative adversarial networks,we find through experiments that it is a more reliable way to choose the Adam optimization algorithm.In addition,the use of label smoothing can improve the problem of unstable training.Secondly,in order to further extend the applicatio n field of GAN,we use Layout GAN to study the layout generation problem in graphic design.Specifically,we first theoretically analyze generator,relation-based discriminator and wireframe rendering discriminator of Layout GAN,then design a network struct ure that conforms to the Layout GAN’s idea,and experiment with the MINST dataset to verify the network structure,experimental results show that wireframe rendering discriminator is better than relation-based discriminator,then only use Layout GAN with wireframe rendering discriminator to experiment on the SUNCG dataset,and finally realize the automatic generation of graphic layout.