Abstract
Image translation has become widely used everywhere and it has been used for generating new images from existing images such as generating avatars, creating fake images, translation of images from one type to another and many more using various machine learning algorithms. One such machine learning technique for image translation is GAN (Generative Adversarial Networks). The classic GANs does the job of translation the images of a domain U to domain V, and the conditional GANs does the same but basing on a condition. These GANs are able to tell the discrimination between the objects accurately when training labeled data sets, and cannot perform well when training unlabeled data sets. So, we need a GAN that can train the unlabeled data sets well and able to translate the images of the domain U to domain V and vice-versa so that we can calculate the realness or fakeness of the image translated taking the original image as reference. For this purpose, we research dual GANs, which combine a primal GAN with another GAN. The inversion of the task of the primal GAN, which is comparable to the traditional GAN, is learnt by dual GAN,
Keyword
Generative Adversarial Networks, Dual GANs, Classic GANS, Machine Learning, Generator, Discriminator, Image translation, Neural networks, Conditional GANs, Standardization
PDF Download (click here)
|