SRGAN

Super Resolution GAN.

This is a re-implementation using Pytorch of SRGAN models based on original paper Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network.

Requirements

• Pytorch
• tqdm
• easydict
• wandb

Datasets

Train Dataset

The train datasets is ImageNet Large Scale Visual Recognition Challenge 2012 (ILSVRC2012). The paper randomly chose 350 000 samples from ImageNet dataset for training process.

Test Dataset

In this code, I use Set5 dataset for val and Set5, Set14, and BSD100 for evaluation. You can download dataset here as refered to this github.

Usage

Train

Setting for training and evaluating: config/config.yaml. Download training dataset and modify train:dataset:data_dir: <path/to/train/dataset>

python src/main.py

Evaluate

python src/eval.py

Result

SRGAN was evaluated on these dataset: BSD100, Set14, Set5. With scale x4: from left to right: low resolution, resolved resolution (SRGAN), high resolution

Set5

• psnr: 30.7253, ssim: 0.9823

Set14

• psnr: 20.3099, ssim: 0.9457

• psnr: 20.9446, ssim: 0.9769

BDS100

• psnr: 31.7421, ssim: 0.9930

• psnr: 29.6615, ssim: 0.9960

• psnr: 23.4305, ssim: 0.9802