| Title |
U-Net-Based Generative Adversarial Network |
| Authors |
박승(Seung Park) ; 신용구(Yong-Goo Shin) |
| DOI |
https://doi.org/10.5573/ieie.2021.58.5.61 |
| Keywords |
Generative adversarial network; U-Net-based generator; Image generation |
| Abstract |
Recently, generative adversarial network (GAN), which generates high-resolution images from an input random-noise vector, has been extensively studied. The traditional generator consisting of multiple residual blocks is trained to generate high-quality image, whereas discriminator is trained to distinguish between real and generated samples. To improve the performance of GAN, conventional methods suggest novel regularization techniques or training strategies rather than modifying the generator architecture. Different with the conventional approaches, we propose a novel generator architecture inspired by U-Net. The proposed method significantly improves the generative performance with marginal training overhead. Extensive experiments with various datasets including CIFAR-10, CelebA, and LSUN show that the proposed method significantly improves the performance of GAN in terms of Frechet inception distance (FID). |