Image Super-Resolution using EDSR & SwinIR

---

Collaborators : 최명헌, 장준혁, 최윤서, 황진우, 조보경

---

Through the Dacon contest dataset, we implemented Image Super-Resolution with our custom model using EDSR and SwinIR as streamlit app.

Overview

Dataset

• Train data : 1640 LR(512×512) & HR(2048×2048) paired images
• Test data : 18 LR images
• Dacon contest link : AI 양재 허브 인공지능 오픈소스 경진대회

Data Preprocessing

We cut the images into 64 patches regardless of resolution and made upscaling model(64×64 ➡ 256×256)

Custom Model

Streamlit

Installation

• Run the command pip install -r requirements.txt to install requirements

Usage

1. Fork this repository and install the requirements as mentioned above
2. Run super_resolution.py with streamlit

streamlit run super_resolution.py

3. Upload your low-resolution image and get high-resolution image

Results

File Description

models

1. data

• crop_attach_image.py : Crop a low-resolution image to 64 patches and attach upscaling images to restore
• customTrain.csv / customValid.csv / test.csv : CSV files which are the lists of train images, valid images and test images

2. modules

• pretrained models : SwinIR pretrained models
• dataset.py : Transform images and build custom dataset
• edsr.py : Customized EDSR model
• swinir.py : Customized SwinIR model
• losses.py : For L1 loss in train process
• model.py : Our own made super-resolution model code

3. others

• inference_dacon.py : Code of super resolution for test set
• main.py : The whole process from building dataset to super resolution
• train.py : Train process code

results

• CustomModel_30.pt : pretrained custom super resolution model

streamlit

• app_funcs.py : Functions that are used in streamlit app
• super_resolution.py : Code to run streamlit app
• uploads : test images

others

• requirements.txt : Required dependencies to run the streamlit app