AnimeTransGAN: Animation Image Super-Resolution Transformer via Deep Generative Adversarial Network

To achieve better visual experiences for watching classic animations displayed on high-end displays, such as UHDTV (ultra high-definition television), a novel deep learning framework designed for animation image super-resolution (SR) is proposed in this paper. To overcome the possible drawbacks that GAN (generative adversarial network)-based SR models may not recover sufficient image details while transformer-based SR models may produce over-blurred/over-sharpened images, we propose a novel GAN-based model for animation image SR, where we integrate the superior detail recovery capability of transformer models for image SR, and the discriminative ability of the U-Net-based discriminator for determining the reality of generated images, denoted by AnimeTransGAN. Our experimental results demonstrate that the proposed AnimeTransGAN model quantitatively and qualitatively achieves better SR performances for animation images, compared with the state-of-the-art methods.