jzrita / ntire2020_sysu-air

NTIRE 2020 Perceptual Extreme Super-Resolution Challenge

A-Fast-Feedback-Network-for-Large-Scale-Image-SR

FFNSR @ NTIRE 2020

This repository is Pytorch code for our proposed FFNSR.

The code is developed by team sysu-AIR, and tested on Ubuntu 18.04 environment (Python 3.6, PyTorch 1.4.0, CUDA 10.1) with 2080Ti GPUs.The details about our proposed FFNSR can be found in our factsheet.

Contents

1. FFNSR
2. Requirements
3. Test
4. Train
5. Result

FFNSR

The sysu-AIR team proposed A Fast Feedback Network for Large Scale Image Super-Resolution. Inspired by SRFBN and IMDN, the proposed FFNSR is still reserved the RNN structure but with a information multi-distillation module (IMDM), which can benefit image SR tasks and accelerate inference speed.

• Number of parameters: 2,099,625

• Average PSNR on validation data: 29.01096 dB

• Average inference time (RTX 2080 Ti) on validation data: 4.35 second

  Note: We selected the best average inference time among three trials

Requirements

• Python 3 (Anaconda is recommended)
• skimage
• imageio
• Pytorch (Pytorch version >=0.4.1 is recommended)
• tqdm
• pandas
• cv2 (pip install opencv-python)

Test

Quick start

1. Clone this repository:

   git clone https://github.com/jzrita/NTIRE2020_sysu-AIR.git

2. Download our pre-trained model from the links below, unzip the models and place them to ./models.

   Click_here_to_download (code: a3mu)

3. CD the folder and install the requirements:

   cd NTIRE2020_sysu-AIR && pip install -r requirements.txt

4. Place the LR pictures to ./picture.

   ./picture/1601.png
   ./picture/1602.png
   ./picture/1603.png
   ...

5. Edit ./options/test/test_SRFBN_example.json for your needs according to ./options/test/README.md.

6. Then run the following commands to test the model:

    python test.py

Train

1. Edit ./options/train/train_SRFBN.json for your needs according to ./options/train/README.md.

2. Run command to train the model：

   cd NTIRE2020_sysu-AIR
   python train.py

3. You can monitor the training process in ./experiments.

Result

1. Download and view our test result.

   Click_here_to_download (code: fiz9)