/MetaSR-PyTorch

PyTorch implements `Meta-SR: A Magnification-Arbitrary Network for Super-Resolution` paper

Primary LanguagePythonApache License 2.0Apache-2.0

MetaSR-PyTorch

Overview

This repository contains an op-for-op PyTorch reimplementation of Meta-SR: A Magnification-Arbitrary Network for Super-Resolution.

Table of contents

Download weights

Download datasets

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.

How Test and Train

Both training and testing only need to modify the config.py file.

Test

Modify the config.py file.

  • line 31: arch_name change to metasr_rdn.
  • line 39: upscale_factor change to 4.
  • line 47: mode change to test.
  • line 49: exp_name change to test_MetaSR_RDN-DIV2K.
  • line 90: model_weights_path change to ./results/pretrained_models/MetaSR_RDN-DIV2K-8daac205.pth.tar.
python3 test.py

Train MetaSR_RDN model

Modify the config.py file.

  • line 31: arch_name change to metasr_rdn.
  • line 39: upscale_factor change to 4.
  • line 47: mode change to train.
  • line 49: exp_name change to MetaSR_RDN-DIV2K.
python3 train.py

Resume train MetaSR_RDN model

Modify the config.py file.

  • line 31: arch_name change to metasr_rdn.
  • line 39: upscale_factor change to 4.
  • line 47: mode change to train.
  • line 49: exp_name change to MetaSR_RDN-DIV2K.
  • line 67: resume_model_weights_path change to ./samples/MetaSR_RDN-DIV2K/epoch_xxx.pth.tar.
python3 train.py

Result

Source of original paper results: https://arxiv.org/pdf/1903.00875v4.pdf

In the following table, the value in () indicates the result of the project, and - indicates no test.

Model Dataset Scale PSNR
MetaSR_RDN Set14 2.0 32.35(32.37)
MetaSR_RDN Set14 2.5 30.45(29.85)
MetaSR_RDN Set14 3.0 29.30(27.95)
MetaSR_RDN Set14 3.5 28.32(27.29)
MetaSR_RDN Set14 4.0 27.75(26.55) 
# Download `MetaSR_RDN-DIV2K-8daac205.pth.tar` weights to `./results/pretrained_models`
# More detail see `README.md<Download weights>`
python3 ./inference.py

Input:

Output:

Build `metasr_rdn` model successfully.
Load `metasr_rdn` model weights `./results/pretrained_models/MetaSR_RDN-DIV2K-8daac205.pth.tar` successfully.
SR image save to `./figure/comic_sr.png`

Credit

Meta-SR: A Magnification-Arbitrary Network for Super-Resolution

Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, Jian Sun

Abstract
Recent research on super-resolution has achieved great success due to the development of deep convolutional neural networks (DCNNs). However, super-resolution of arbitrary scale factor has been ignored for a long time. Most previous researchers regard super-resolution of different scale factors as independent tasks. They train a specific model for each scale factor which is inefficient in computing, and prior work only take the super-resolution of several integer scale factors into consideration. In this work, we propose a novel method called Meta-SR to firstly solve super-resolution of arbitrary scale factor (including non-integer scale factors) with a single model. In our Meta-SR, the Meta-Upscale Module is proposed to replace the traditional upscale module. For arbitrary scale factor, the Meta-Upscale Module dynamically predicts the weights of the upscale filters by taking the scale factor as input and use these weights to generate the HR image of arbitrary size. For any low-resolution image, our Meta-SR can continuously zoom in it with arbitrary scale factor by only using a single model. We evaluated the proposed method through extensive experiments on widely used benchmark datasets on single image super-resolution. The experimental results show the superiority of our Meta-Upscale.

[Code(PyTorch)] [Paper]

@article{hu2019meta,
  title={Meta-SR: A Magnification-Arbitrary Network for Super-Resolution},
  author={Hu, Xuecai and Mu, Haoyuan and Zhang, Xiangyu and Wang, Zilei  and Tan, Tieniu and Sun, Jian},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
  year={2019}
}