Research And Application Of GAN In Image Translation

Generation antagonistic network(GAN)was proposed in 2014 to help generate realistic visual images,which has become one of the hottest research subjects in deep learning in recent years.In the field of image translation,although GAN is more effective than traditional methods,it has some disadvantages such as difficulty in training,unstable network and difficulty in convergence.In view of these disadvantages,a large number of researchers have made improvements on GAN,and it becomes very meaningful to improve GAN's performance in the field of image translation.In this paper,GAN is applied in the field of image translation,a new framework is proposed,and the image translation task based on supervised learning and unsupervised learning is completed.First,put forward a kind of new generation against network framework,based on the condition of boundary equilibrium generated against network(CBEGAN),a clear picture of the model can be generated according to the conditions specified,is the foundation of image translation framework,the experimental results show that the method is compared with other supervision class generation model can be more easy to use the network to achieve faster convergence speed and the ability to generate a better quality and diversity of images.Secondly,the model is improved and a supervised image translation algorithm is proposed by constructing an encoder to complete the image translation of data.Finally,the unsupervised image translation task is transformed into supervised image translation task to realize the image translation task of the model between face photos and sketch images under unpaired data.The experimental results show that the image translation effect of paired data and unpaired data is better than that of other models.