Skip to content

EnhanceNet: Single Image Super-Resolution Through Automated Texture Synthesis (official repository)

License

Notifications You must be signed in to change notification settings

msmsajjadi/EnhanceNet-Code

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

EnhanceNet

This is a pre-trained reference implementation of ENet-PAT from "EnhanceNet: Single Image Super-Resolution through Automated Texture Synthesis" for a magnification ratio of 4.

It comes with a script which takes care of installing all necessary packages inside a virtual environment, i.e. all installations take place inside the folder and can be simply uninstalled by deleting the folder. Advanced users may install the packages directly (see FAQ section).

If you use this code as part of a publication, please cite:

@inproceedings{enhancenet,
  title={{EnhanceNet: Single Image Super-Resolution through Automated Texture Synthesis}},
  author={Sajjadi, Mehdi S. M. and Sch{\"o}lkopf, Bernhard and Hirsch, Michael},
  booktitle={Computer Vision (ICCV), 2017 IEEE International Conference on},
  pages={4501--4510},
  year={2017},
  organization={IEEE},
  url={https://arxiv.org/abs/1612.07919/}
}

How-To

Upscale images 4x

  • Copy all high-resolution images to the input directory.
  • Enter the following into a terminal window: ./run.sh

The images will be first downscaled and then upscaled with EnhanceNet.

Requirements:

Notes

  • All images in the input folder are downscaled 4x and then upscaled via bicubic interpolation and EnhanceNet-PAT.

  • Existing files in the output directory will not be replaced.

  • The first run may take a while, since all necessary packages are installed. Subsequent runs are much faster and do not necessitate an internet connection.

  • For compatibility reasons, the CPU version of TensorFlow is used and the TF computation graph is rebuilt for each image, so this reference implementation does not reflect the runtime performance of our model and is not suitable for runtime benchmarks.

FAQ

  • Upon calling ./run.sh, the message Permission denied appears.

    Solution: Enter chmod 544 run.sh enhancenet.py, then retry the run command.

  • I already have TensorFlow and/or don't want to run the virtualenv script.

    Solution: Install the necessary packages (see run.sh), then run python enhancenet.py

  • I have CUDA and want this to be faster for large images.

    Solution: In run.sh, add "-gpu" to the line with tensorflow: tensorflow-gpu=0.12.0

  • I don't want to downscale the images before upscaling, I want the actual upscaled version of 4x the size of the input images.

    Solution: In enhancenet.py, set the scaling factor to 1:

    imgs = loadimg('input/'+fn, 1)

    Please note that the model was specifically trained on input images downsampled with PIL, i.e. it won't perform as well on images downscaled with another method. Furthermore, it is not suitable for upscaling noisy or compressed images, since artifacts in the input will be heavily amplified. For comparisons on such images, the model needs to be trained on such a dataset.

  • I get the error message std::bad_alloc.

    Solution: You are likely running out of memory! Please note that even 1000x1000 input images need a lot of RAM (when the input is not downsampled).

  • Something goes wrong and I am left with an error message.

    Solution: Please email us, see below (even if you solved it, so we can add it to the FAQ).

For any questions, comments or help to get it to run, please don't hesitate to mail us: msajjadi@tue.mpg.de

Releases

No releases published

Packages

No packages published