README_back.md

SRGAN-PyTorch

Overview

This repository contains an op-for-op PyTorch reimplementation of Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network .

Table of contents

• SRGAN-PyTorch
  • Overview
  • Table of contents
  • Download weights
  • Download dataset
    • Download train dataset
    • Download valid dataset
  • Test
  • Train
    • Train SRResNet model
    • Train SRGAN model
  • Result
  • Contributing
  • Credit
    • Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network

Download weights

• Google Driver
• Baidu Driver access:llot

Download dataset

Download train dataset

ImageNet

• Image format

  • Google Driver
  • Baidu Driver access: llot

• LMDB format (train)

  • Google Driver
  • Baidu Driver access: llot

• LMDB format (valid)

  • Google Driver
  • Baidu Driver access: llot

Download valid dataset

Set5

• Image format
  • Google Driver
  • Baidu Driver access:llot

Set14

• Image format
  • Google Driver
  • Baidu Driver access:llot

BSD100

• Image format
  • Google Driver
  • Baidu Driver access:llot

BSD200

• Image format
  • Google Driver
  • Baidu Driver access:llot

Test

Modify the contents of the file as follows.

• line 28: upscale_factor change to the magnification you need to enlarge.
• line 30: mode change Set to valid mode.
• line 113: model_path change weight address after training.

Train

Modify the contents of the file as follows.

• line 28: upscale_factor change to the magnification you need to enlarge.
• line 30: mode change Set to train mode.

If you want to load weights that you've trained before, modify the contents of the file as follows.

Train SRResNet model

• line 47: resume change to True.
• line 48: strict Transfer learning is set to False, incremental learning is set to True.
• line 49: start_epoch change number of training iterations in the previous round.
• line 50: resume_weight the weight address that needs to be loaded.

Train SRGAN model

• line 75: resume change to True.
• line 76: strict Transfer learning is set to False, incremental learning is set to True.
• line 77: start_epoch change number of training iterations in the previous round.
• line 78: resume_d_weight the discriminator weight address that needs to be loaded.
• line 79: resume_g_weight the generator weight address that needs to be loaded.

Result

Source 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.

Dataset	Scale	SRResNet (PSNR)	SRGAN (PSNR)
Set5	4	32.05(31.80)	29.40(28.88)
Set14	4	28.49(28.20)	26.02(25.70)

Low resolution / Recovered High Resolution / Ground Truth

Contributing

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!

Credit

Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network

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.

[Paper]

@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}
}