RIS-GAN: Self-Supervised GANs via Recovering Initial State of Subimages.

• RIS-GAN is proposed for the self-supervised learning of generative adversarial networks. • RIS-GAN greatly improves the quality of extracted features and generated images. • RIS-GAN brings the network more stable training property. Self-supervised methods play a vital role in representation learning without the need for annotated data and learn feature representations where data itself provides supervision. However, it is still challenging for the discriminator to learn reliable representations. To address this issue, we propose the self-supervised GANs via recovering initial state of subimages (RIS-GAN). Recovering initial state (RIS) task requires the network to recover the original state information of each sub-part of a composite image. Specifically, the network should predict the rotation angle and the initial position of subimages, and also predict which original image each subimage belongs to. To recover these information, the network should capture intra-image and inter-image information and extract high-similarity features between those patches belonging to the same original images. Besides, we propose local-discriminator, to improve the ability of the discriminator to judge global features from local patch features. To accomplish the above task, our proposed model should extract multi-level information to improve the understanding of image structure and content, and the ability to mine valuable information from image patches. We experimentally demonstrate the state-of-the-art performance of our RIS-GAN with respect to other self-supervised GAN methods on CIFAR-10, ImageNet 32 × 32, CelebA, and STL-10 datasets. [ABSTRACT FROM AUTHOR]