Super resolution image

SRResnet & SRGAN implementations.

Dataset

I used VOC2012 with 17k pair of images (LR, HR) to train models. (LR, HR) with size (22x22) upto (88x88) with scale factor = 4.

Then test phase with 3 benchmark datasets: BDS100, Set14, Set5.

Result

Here is the result after 50 epochs with default setting.

Note: PSNR/SSIM

	BDS100	SET14	SET5
SRResnet	26.68/0.82	30.87/0.94	27.46/0.86
SRGAN	26.07/0.81	29.36/0.92	26.49/0.85

Training

• Git clone this repo
• Install requirements
• Create checkpoint folder inside
• Prepocessing data
• Train and Inference

You can change the setting directly in train_SRResnet.py or train_SRGAN.py.

Run with default setting.

python train_SRResnet.py
python train_SRGAN.py

To inference

python inference.py

Prepocessing

I compressed dataset in .pkl file for training on google colab. Check notebook for more details. Pkl saving format: List of np array.

Download pkl file with demo [1000/300]:

Report

Report