This repository contains an op-for-op PyTorch reimplementation of Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network.
Contains DIV2K, DIV8K, Flickr2K, OST, T91, Set5, Set14, BSDS100 and BSDS200, etc.
Please refer to README.md in the data directory for the method of making a dataset.
Both training and testing only need to modify the srresnet_config.py file and srgan_config.py file.
Modify the srgan_config.py file.
- line 32:
g_arch_namechange tosrresnet_x4. - line 39:
upscale_factorchange to4. - line 41:
modechange totest. - line 43:
exp_namechange toSRGAN_x4-DIV2K. - line 96:
g_model_weights_pathchange to./results/pretrained_models/SRGAN_x4-ImageNet-8c4a7569.pth.tar.
python3 test.pyModify the srresnet_config.py file.
- line 32:
g_arch_namechange tosrresnet_x4. - line 39:
upscale_factorchange to4. - line 41:
modechange totrain. - line 43:
exp_namechange toSRResNet_x4-DIV2K.
python3 train_srresnet.pyModify the srresnet_config.py file.
- line 32:
g_arch_namechange tosrresnet_x4. - line 39:
upscale_factorchange to4. - line 41:
modechange totrain. - line 43:
exp_namechange toSRResNet_x4-DIV2K. - line 59:
resume_g_model_weights_pathchange to./samples/SRGAN_x4-DIV2K/g_epoch_xxx.pth.tar.
python3 train_srresnet.py- line 31:
d_arch_namechange todiscriminator. - line 32:
g_arch_namechange tosrresnet_x4. - line 39:
upscale_factorchange to4. - line 41:
modechange totrain. - line 43:
exp_namechange toSRGAN_x4-DIV2K. - line 58:
pretrained_g_model_weights_pathchange to./results/SRResNet_x4-DIV2K/g_last.pth.tar.
python3 train_srgan.py- line 31:
d_arch_namechange todiscriminator. - line 32:
g_arch_namechange tosrresnet_x4. - line 39:
upscale_factorchange to4. - line 41:
modechange totrain. - line 43:
exp_namechange toSRGAN_x4-DIV2K. - line 61:
resume_d_model_weights_pathchange to./samples/SRGAN_x4-DIV2K/d_epoch_xxx.pth.tar. - line 62:
resume_g_model_weights_pathchange to./samples/SRGAN_x4-DIV2K/g_epoch_xxx.pth.tar.
python3 train_srgan.pySource 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(32.14) | 29.40(30.64) |
| SSIM | 4 | 0.9019(0.8954) | 0.8472(0.8642) |
| 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) |
# Download `SRGAN_x4-ImageNet-8c4a7569.pth.tar` weights to `./results/pretrained_models`
# More detail see `README.md<Download weights>`
python3 ./inference.pyInput:
Output:
Build `srresnet_x4` model successfully.
Load `srresnet_x4` model weights `./results/pretrained_models/SRGAN_x4-ImageNet-8c4a7569.pth.tar` successfully.
SR image save to `./figure/comic_sr.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}
}