Exploring real-time super-resolution generative adversarial networks

摘要

Image super-resolution is an essential technology for improving user quality of experience of internet videos. As the state-of-the-art deep learning-based super-resolution technology, the enhanced super-resolution generative adversarial networks (ESRGAN) has the best performance in the perceptual quality of reconstructed images, and the efficient sub-pixel convolutional neural network (ESPCN) has the best real-time performance. This work proposes real-time super-resolution generative adversarial network (RTSRGAN). RTSRGAN takes the advantages of ESRGAN and ESPCN so as to simultaneously satisfy the demands on the real-time performance and the resulting pleasant artifacts of super-resolution at the client side. Our experimental studies demonstrate our proposed RTSRGAN can be used for super-resolution at the client side to enhance the real-time performance as well as ensure the image perceptual quality. We also find that RTSRGAN is suitable for restoring images with regularly changing texture features without requiring training for individual image categories.