Image Super Resolution

Super resolution (SR) is the process of taking an input of a low resolution (LR) and upscaling it to that of a high resolution. This program is the implementation of the paper Enhanced Deep Residual Networks for Single Image Super-Resolution by Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee.

Hardware

• Intel(R) Core(TM) i5-9600K CPU @ 3.70GHz
• NVIDIA GeForce RTX 2080 Ti

Environment

• Microsoft win10
• Python 3.7.3
• Pytorch 1.7.0
• CUDA 10.2

Install Packages

• install pytorch from https://pytorch.org/get-started/locally/
• install dependencies

pip install -r requirements.txt

Data Preparation

Download the given dataset from Google Drive and put it into src folder. Then run python convert.py to generate the corresponding images and folders.

src
├── DIV2K
│   ├── DIV2K_train_HR
│   └── DIV2K_train_LR_bicubic
│       ├── X2
│       └── X3
└── hw4
    ├── testing_lr_images
    └── training_hr_images

Training

Execute following instruction to train the model:

python main.py --model EDSR --scale 3 --patch_size 30 --save edsr_baseline_x3 --reset --dir_data . --lr 0.00005

※ get more info by python main.py -h

Testing

Execute following instruction to test images:

python main.py --data_test Demo --dir_demo ./hw4/testing_lr_images/testing_lr_images --scale 3 --pre_train ../experiment/edsr_baseline_x3/model/model_best.pt --test_only --save_results --save HW4_result