Home

Awesome

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

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.

MethodsScalePSNRSSIMLPIPS
Bicubic232.670.8870.201
EDSR234.240.9080.155
RCAN234.340.9080.158
CDC(ours)234.450.9100.146
Bicubic331.500.8350.362
EDSR332.930.8760.241
RCAN333.030.8760.241
CDC(ours)333.060.8760.244
Bicubic430.560.8200.438
EDSR432.030.8550.307
RCAN431.850.8570.305
CDC(ours)432.420.8610.300

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}
}