/Component-Divide-and-Conquer-for-Real-World-Image-Super-Resolution

Primary LanguagePython

Component Divide-and-Conquer for Real-World Image Super-Resolution(CDC)

This repository is an official PyTorch implementation of the paper "Component Divide-and-Conquer for Real-World Image Super-Resolution " from ECCV 2020. [PDF]

We provide full training and testing codes, pre-trained models and the large-scale dataset used in our paper. You can train your model from scratch, or use a pre-trained model to enlarge your images.

Code

Dependencies

  • Python 3.6
  • PyTorch >= 1.1.0
  • numpy
  • cv2
  • skimage
  • tqdm

Quick Start

Clone this github repo.

git clone https://github.com/xiezw5/Component-Divide-and-Conquer-for-Real-World-Image-Super-Resolution
cd Component-Divide-and-Conquer-for-Real-World-Image-Super-Resolution/CDC

Training

  1. Download our dataset and unpack them to any place you want. Then, change the dataroot and test_dataroot argument in ./options/realSR_HGSR_MSHR.py to the place where images are located.
  2. Run CDC_train_test.py using script file train_pc.sh.
sh ./train_pc.sh cdc_x4 ./CDC_train_test.py ./options/realSR_HGSR_MSHR.py 1
  1. You can find the results in ./experiments/CDC-X4 if the exp_name argument in ./options/realSR_HGSR_MSHR.py is CDC-X4

Testing

  1. Download our pre-trained models to ./models folder or use your pre-trained models
  2. Change the test_dataroot argument in CDC_test.py to the place where images are located
  3. Run CDC_test.py using script file test_models_pc.sh.
sh test_models_pc.sh cdc_x4_test ./CDC_test.py ./models/HGSR-MHR_X4_SubRegion_GW_283.pth 1
  1. You can find the enlarged images in ./results folder

Pretrained models

  1. 2X Models
  2. 3X Models
  3. 4X Models

The above provided models are both trained on our dataset with our gradient-weighted loss.

Dataset

Please download our dataset from Google Drive or Baidu Drive. The verification code is osiy. There are 31970 192×192 patches cropped for training and 93 image pairs for testing.

Methods Scale PSNR SSIM LPIPS
Bicubic 2 32.67 0.887 0.201
EDSR 2 34.24 0.908 0.155
RCAN 2 34.34 0.908 0.158
CDC(ours) 2 34.45 0.910 0.146
Bicubic 3 31.50 0.835 0.362
EDSR 3 32.93 0.876 0.241
RCAN 3 33.03 0.876 0.241
CDC(ours) 3 33.06 0.876 0.244
Bicubic 4 30.56 0.820 0.438
EDSR 4 32.03 0.855 0.307
RCAN 4 31.85 0.857 0.305
CDC(ours) 4 32.42 0.861 0.300

Citation

If you find our work useful in your research or publication, please cite:

@InProceedings{wei2020cdc,
  author = {Pengxu Wei, Ziwei Xie, Hannan Lu, ZongYuan Zhan, Qixiang Ye, Wangmeng Zuo, Liang Lin},
  title = {Component Divide-and-Conquer for Real-World Image Super-Resolution},
  booktitle = {Proceedings of the European Conference on Computer Vision},
  year = {2020}
}