This repository contains an op-for-op PyTorch reimplementation of Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network.
Download all available model weights.
# Download `SRGAN_x4-SRGAN_ImageNet.pth.tar` weights to `./results/pretrained_models`
$ bash ./scripts/download_weights.sh SRGAN_x4-SRGAN_ImageNet
# Download `SRResNet_x4-SRGAN_ImageNet.pth.tar` weights to `./results/pretrained_models`
$ bash ./scripts/download_weights.sh SRResNet_x4-SRGAN_ImageNet
# Download `DiscriminatorForVGG_x4-SRGAN_ImageNet.pth.tar` weights to `./results/pretrained_models`
$ bash ./scripts/download_weights.sh DiscriminatorForVGG_x4-SRGAN_ImageNetThese train images are randomly selected from the verification part of the ImageNet2012 classification dataset.
$ bash ./scripts/download_datasets.sh SRGAN_ImageNetIt is convenient to download some commonly used test data sets here.
$ bash ./scripts/download_datasets.sh Set5Both training and testing only need to modify yaml file.
Set5 is used as the test benchmark in the project, and you can modify it by yourself.
If you need to test the effect of the model, download the test dataset.
$ bash ./scripts/download_datasets.sh Set5$ python3 test.py --config_path ./configs/test/SRGAN_x4-SRGAN_ImageNet-Set5.yaml$ python3 test.py --config_path ./configs/test/SRResNet_x4-SRGAN_ImageNet-Set5.yamlFirst, the dataset image is split into several small images to reduce IO and keep the batch image size uniform.
$ python3 ./scripts/split_images.pyThen, run the following commands to train the model
$ python3 train_net.py --config_path ./configs/train/SRResNet_x4-SRGAN_ImageNet.yamlModify the ./configs/train/SRResNet_x4-SRGAN_ImageNet.yaml file.
- line 33:
RESUMED_G_MODELchange to./samples/SRResNet_x4-SRGAN_ImageNet/g_epoch_xxx.pth.tar.
$ python3 train_net.py --config_path ./configs/train/SRResNet_x4-SRGAN_ImageNet.yaml$ python3 train_gan.py --config_path ./configs/train/SRGAN_x4-SRGAN_ImageNet.yamlModify the ./configs/train/SRGAN_x4-SRGAN_ImageNet.yaml file.
- line 38:
PRETRAINED_G_MODELchange to./results/SRResNet_x4-SRGAN_ImageNet/g_last.pth.tar. - line 40:
RESUMED_G_MODELchange to./samples/SRGAN_x4-SRGAN_ImageNet/g_epoch_xxx.pth.tar. - line 41:
RESUMED_D_MODELchange to./samples/SRGAN_x4-SRGAN_ImageNet/d_epoch_xxx.pth.tar.
$ python3 train_gan.py --config_path ./configs/train/SRGAN_x4-SRGAN_ImageNet.yamlSource of original paper results: https://arxiv.org/pdf/1609.04802v5.pdf
In the following table, the psnr value in () indicates the result of the project, and - indicates no test.
| Set5 | Scale | SRResNet | SRGAN |
|---|---|---|---|
| PSNR | 4 | 32.05(32.16) | 29.40(30.67) |
| SSIM | 4 | 0.9019(0.8938) | 0.8472(0.8627) |
| Set14 | Scale | SRResNet | SRGAN |
|---|---|---|---|
| PSNR | 4 | 28.49(28.57) | 26.02(27.12) |
| SSIM | 4 | 0.8184(0.7815) | 0.7397(0.7321) |
| BSD100 | Scale | SRResNet | SRGAN |
|---|---|---|---|
| PSNR | 4 | 27.58(27.56) | 25.16(26.22) |
| SSIM | 4 | 0.7620(0.7367) | 0.6688(0.6867) |
# If you do not train the model yourself, you can download the model weights and test them.
$ bash ./scripts/download_weights.sh SRGAN_x4-SRGAN_ImageNet
$ python3 ./inference.pyInput:
Output:
Build `srresnet_x4` model successfully.
Load `srresnet_x4` model weights `SRGAN-PyTorch/results/pretrained_models/SRGAN_x4-SRGAN_ImageNet.pth.tar` successfully.
SR image save to `./figure/sr_comic.png`
If you find a bug, create a GitHub issue, or even better, submit a pull request. Similarly, if you have questions, simply post them as GitHub issues.
I look forward to seeing what the community does with these models!
Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan
Wang, Wenzhe Shi
Abstract
Despite the breakthroughs in accuracy and speed of single image super-resolution using faster and deeper convolutional neural networks, one central
problem remains largely unsolved: how do we recover the finer texture details when we super-resolve at large upscaling factors? The behavior of
optimization-based super-resolution methods is principally driven by the choice of the objective function. Recent work has largely focused on
minimizing the mean squared reconstruction error. The resulting estimates have high peak signal-to-noise ratios, but they are often lacking
high-frequency details and are perceptually unsatisfying in the sense that they fail to match the fidelity expected at the higher resolution. In this
paper, we present SRGAN, a generative adversarial network (GAN) for image super-resolution (SR). To our knowledge, it is the first framework capable
of inferring photo-realistic natural images for 4x upscaling factors. To achieve this, we propose a perceptual loss function which consists of an
adversarial loss and a content loss. The adversarial loss pushes our solution to the natural image manifold using a discriminator network that is
trained to differentiate between the super-resolved images and original photo-realistic images. In addition, we use a content loss motivated by
perceptual similarity instead of similarity in pixel space. Our deep residual network is able to recover photo-realistic textures from heavily
downsampled images on public benchmarks. An extensive mean-opinion-score (MOS) test shows hugely significant gains in perceptual quality using SRGAN.
The MOS scores obtained with SRGAN are closer to those of the original high-resolution images than to those obtained with any state-of-the-art method.
@InProceedings{srgan,
author = {Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, Wenzhe Shi},
title = {Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network},
booktitle = {arXiv},
year = {2016}
}