科研之友
微信
新浪微博
Facebook
分享链接
摘要
Generative adversarial networks (GANs) have achieved remarkable success in image generation, especially training conditional GANs for deriving reliable representations. However, the main downside of conditional GANs is the requirement of labeled data. Using self-supervision information can meet such needs, but the challenges remain for discovering more reliable self-supervised signals and ways to couple different signals to describe characteristics of the training data more precisely. In this paper, we propose a novel self-supervised learning approach to automatically generating pseudo labels for autoencoder-based GANs. Specifically, we blend the input images and the corresponding reconstructed results to produce transformed samples controlled by the blend ratio. Then, an additional classifier attached to the discriminator needs to distinguish the ratio of real images from the transformed samples to derive meaningful representations. Next, we enhance GANs with multiple self-supervision guidances by two different means to further improve the capacity of the discriminator. One merges multiple supervision signals and requires the classifier to predict the mixed probability, whereas the other one utilizes these signals independently. In experiments, we evaluate the quality of the generated image and the learned representation using three datasets. Empirical results prove the effectiveness of our methods on both image synthesis and representation learning.
