StarGAN Improvements: Automatically Generated Labels

Recent studies have shown that image-to-image translation methods have achieved remarkable success in two domains. In the context of the increasingly hot face generation technology, the StarGAN is one of the most effect generation methods. The existing StarGAN can achieve multi-domain learning through the mapping relationship between image generators, it solves the limited scalability and robustness of existing methods when dealing with more than two domains. However, because StarGAN uses one-hot encoding and latent code to input to the model, that is, the label is a fixed tensor, which greatly increases the burden of manpower and time. In this paper, to get around this limitation of traditional StarGAN, a ResNet based multi-label method is designed to automatically generate labels for StarGAN. Our method can effectively augment the training set without adding labor.