Staged Generative Adversarial Networks With Adversarial-Boundary

Generative Adversarial Networks ( GANs) provide a novel way to learn disentangled representations. However, it is still challenging for them to generate convincing images. In this paper we introduce a novel Adversarial-Boundary Staged Generative Adversarial Networks ( ABS-GAN) to generate more realistic images. ABS-GAN improves image quality from two aspects. On one hand, the complete training process is separated into two stages. The Stage-I generator is trained for decreasing the Earth-Mover distance between real and generated distributions. The Stage-II generator aims at explicitly reducing the distance further based on the Stage-I generator. On the other hand, the discriminator is treated as a projector from image to scalar. The discriminator tries to make the boundary between real and generated distributions clear in scalar space. Thus the generator synthesizes more realistic images, thanks to the extra adversarial boundary information. We conduct experiments on real-world datasets ( CIFAR-10, STL-10, CelebA) to show the performance of our ABS-GAN. Comparisons with baseline model on benchmark datasets demonstrate that the proposed method achieves excellent improvement in producing convincing images in a simple way.