SRDGAN: learning the noise prior for Super Resolution with Dual Generative Adversarial NetworksDownload PDF

25 Sep 2019 (modified: 24 Dec 2019)ICLR 2020 Conference Blind SubmissionReaders: Everyone
  • Original Pdf: pdf
  • Keywords: Super Resolution GAN Denoise
  • Abstract: Single Image Super Resolution (SISR) is the task of producing a high resolution (HR) image from a given low-resolution (LR) image. It is a well researched prob- lem with extensive commercial applications like digital camera, video compres- sion, medical imaging, etc. Most recent super resolution works focus on the fea- ture learning architecture, like Chao Dong (2016); Dong et al. (2016); Wang et al. (2018b); Ledig et al. (2017). However, these works suffer from the following chal- lenges: (1) The low-resolution (LR) training images are artificially synthesized us- ing HR images with bicubic downsampling, which have much more information than real demosaic-upscaled images. The mismatch between training and realistic mobile data heavily blocks the effect on practical SR problem. (2) These methods cannot effectively handle the blind distortions during super resolution in practical applications. In this work, an end-to-end novel framework, including high-to-low network and low-to-high network, is proposed to solve the above problems with dual Generative Adversarial Networks (GAN). First, the above mismatch prob- lems are well explored with the high-to-low network, where clear high-resolution image and the corresponding realistic low-resolution image pairs can be gener- ated. With high-to-low network, a large-scale General Mobile Super Resolution Dataset, GMSR, is proposed, which can be utilized for training or as a bench- mark for super resolution methods. Second, an effective low-to-high network (super resolution network) is proposed in the framework. Benefiting from the GMSR dataset and novel training strategies, the proposed super resolution model can effectively handle detail recovery and denoising at the same time.
4 Replies