ESinGAN: Enhanced Single-Image GAN Using Pixel Attention Mechanism for Image Super-Resolution

Recently, the SinGAN model emerged, and it was famous for generating from a single image. The SinGAN model achieves superior performance to other advanced models for image super-resolution task (trained on a dataset or a single image). However, SinGAN does not consider the importance of feature pixels on the feature map. In this paper, we propose an Enhanced SinGAN model (ESinGAN), an unconditional generative model that can improve the defects of SinGAN using the Pixel Attention mechanism. To evaluate the proposed ESinGAN model's performance, we carry out the experiments on a benchmark dataset. Through the experiments, we have achieved better performance than SinGAN and proved the effectiveness of the proposed method. Not only is the image significantly improved visually, but the PSNR and SSIM values of the model are also considerably increased. Besides, ESinGAN runs as fast as SinGAN under the same experimental environment.