Ruofan Zhou Sabine Süsstrunk

Image and Visual Representation Lab
École Polytechnique Fédérale de Lausanne

Abstract

Deep convolutional neural networks (CNNs) have achieved great success in single image super-resolution (SISR) on bicubic down-sampled low-resolution (LR) images. However, their performance is still limited when applying to real images. As the bicubic blur-kernels assumed in these super-resolution approaches deviate from real camera-blur, these methods could visually even perform worse than simple SISR methods such as bicubic interpolation. To improve the generalization capability and robustness of deep super-resolution networks on real photographs, we present a kernel modeling super-resolution network (KMSR) by incorporating blur-kernel modeling in the network. Our proposed KMSR consists of two stages: first building a pool of realistic blur-kernels with a generative adversarial network (GAN), then training a super-resolution network with a dataset constructed from the kernels generated in the first stage. Our extensive experiments demonstrate the effectiveness of our approach in blind image super-resolution on images with unknown blur-kernels.

Materials

[ Paper ] [ Supplementary Material (zip) ] [ Supplementary Material (website) ] [ Poster ]
[ Code (github) ]
[ Phychovisual Experiment ]

Results on X2

Citation

Ruofan Zhou, Sabine Süsstrunk
Kernel Modeling Super-Resolution on Real Low-Resolution Images
IEEE Conference on International Conference on Computer Vision (ICCV), 2019.

          
  @INPROCEEDINGS{7532767, 
    author={R. Zhou and S. Süsstrunk}, 
    booktitle={2019 IEEE Conference on International Conference on Computer Vision (ICCV)}, 
    title={Kernel Modeling Super-Resolution on Real Low-Resolution Images}, 
    year={2019},}

Acknowledgements

The authors thank the anonymous reviewers for their useful comments and suggestions.