.png){kind=icon}
.png){kind=icon}
Return
Image super-resolution using only low-resolution images
.png){kind=icon}
PRE
AI
DOI：10.1007/s00371-022-02646-4.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Currently, most image super-resolution methods use the paired training dataset which is difficult to collect. So bicubic mode is often used to generate paired low-resolution and high-resolution images, but this does not match the actual degradation mode which is often unknown. This results in deteriorated performance of these methods. Considering that only low-resolution images can actually be acquired, we propose a novel model architecture named Cycle-SRGAN which uses only low-resolution images to achieve image super-resolution. The Cycle-SRGAN consists of three components: an up-sampler, a down-sampler and a discriminator. The up-sampler generates high-resolution images, and the down-sampler learns the degradation mode of the images. The discriminator is adopted to help learn the downscaling mode. And we apply two cycle losses, an adversarial loss and a regularization term to implement training. Our method can realize downscaling kernel estimation and image super-resolution at the meantime. In the end, the experimental results indicate the images generated by Cycle-SRGAN own higher image quality than some other methods. And our method has good stability and generalization performance for unknown degradation modes.
Keywords:
Image super-resolution
Unknown degradation mode
Low-resolution images
Image quality
Journal
IF:
2.9
Papers:
4.6K
Citations:
6.5K