muzafferkal / ldrn

LDRN

This repository is implementation of the Learning Local Distribution for Extremely Efficient Single-Image Super-Resolution

Requirements

• PyTorch 1.10.0
• opencv-python 4.5.2.52
• Numpy 1.20.1
• torchvision 0.11.1
• tqdm 4.62.2

Train

The DIV2K, Set5, Set14, B100, Urban100, Manga109 dataset can be downloaded from the links below.

• DIV2K
• Set5
• Set14
• B100
• Urban100
• Manga109

Otherwise,you can use dataProcess.py to create custom dataset and use dataProcess2.py to create GaussianBlur dataset. Get start with the following command

python train.py --line_weight 12 1 --scale 2 --cuda

The pre-training weight file is saved in the weight directory, and you can also start with the following command

python train.py --line_weight 12 1 --scale 2 --pre_train "weight/LDRN_X2_bestpsnr_37.35.pth" --cuda

Results

set5(Bicubic)

Eval.Mat	Scale	LDRN
PSNR/SSIM	2	37.35/0.9812
PSNR/SSIM	3	33.86/0.9666
PSNR/SSIM	4	30.82/0.9349

set5(Gaussian)

Eval.Mat	Scale	LDRN
PSNR/SSIM	2	37.24/0.9809
PSNR/SSIM	3	33.69/0.9658
PSNR/SSIM	4	30.72/0.9345

x3

x4