Introduction

PyTorch implementation of efficient image super-resolution models.

Requirements

torch == 1.8.1
torchmetrics
loguru
tqdm

Supported models

• CARN ¹
• DRCN ²
• DRRN ³
• EDSR ⁴
• ESPCN ⁵
• FSRCNN ⁶
• IDN ⁷
• LapSRN ⁸
• SRCNN ⁹
• SRDenseNet ¹⁰
• VDSR ¹¹

How to use

DDP training (recommend)

CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 main.py

DP training

CUDA_VISIBLE_DEVICES=0,1,2,3 python main.py

Performances and checkpoints

Model	Year	Train on1	Set5		Set14		BSD100
2x			PSNR (paper/my)	SSIM	PSNR	SSIM	PSNR	SSIM
CARN	2018	T+B+D	37.76/37.90	0.9590/0.9605	33.52/33.14	0.9166/0.9152	32.09/32.06	0.8978/0.8985
DRCN	2015	T	37.63/37.85	0.9588/0.9604	33.04/33.22	0.9118/0.916	31.85/32.05	0.8942/0.8982
DRRN	2017	T+B	37.74/37.76	0.9591/0.9599	33.23/33.14	0.9136/0.9149	32.05/31.99	0.8973/0.8974
EDSR	2017	D	37.99/37.90	0.9604/0.9606	33.57/33.22	0.9175/0.9163	32.16/32.10	0.8994/0.899
ESPCN	2016	I+T	n.a./36.85	n.a./0.9559	n.a./32.31	n.a./0.9087	n.a./31.40	n.a./0.8897
FSRCNN	2016	T+G	37.00/37.27	0.9558/0.958	32.63/32.65	0.9088/0.9115	31.53/31.67	0.8920/0.8934
IDN	2018	T+B	37.83/37.84	0.96/0.9604	33.30/33.12	0.9148/0.9155	32.08/32.06	0.8985/0.8985
LapSRN	2017	T+B	37.52/37.59	0.9591/0.9592	32.99/32.96	0.9124/0.9138	31.80/31.89	0.8952/0.8961
SRCNN	2014	I+T	36.66/36.88	0.9542/0.9561	32.45/32.42	0.9067/0.9092	31.36/31.50	0.8879/0.8907
SRDenseNet	2017	I	n.a./37.67	n.a./0.9596	n.a./33.05	n.a./0.9142	n.a./31.93	n.a./0.8967
VDSR	2015	T+B	37.53/37.74	0.9587/0.9598	33.03/33.06	0.9124/0.9145	31.90/31.97	0.8960/0.8973

[¹ Original training dataset, which are short for B (BSD200), D (DIV2K), G (General100), I (ImageNet), T (T91). In my experiments, the training dataset is T + G + B.]

Prepare the dataset

/train
    /T91
    /General100
    /BSD200
/val
    /Set5
    /Set14
    /BSD100

References

Footnotes

1. Fast, Accurate, and Lightweight Super-Resolution with Cascading Residual Network ↩

2. Deeply-Recursive Convolutional Network for Image Super-Resolution ↩

3. Image Super-Resolution via Deep Recursive Residual Network ↩

4. Enhanced Deep Residual Networks for Single Image Super-Resolution ↩

5. Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network ↩

6. Accelerating the Super-Resolution Convolutional Neural Network ↩

7. Fast and Accurate Single Image Super-Resolution via Information Distillation Network ↩

8. Deep Laplacian Pyramid Networks for Fast and Accurate Super-Resolution ↩

9. Image Super-Resolution Using Deep Convolutional Networks ↩

10. Image Super-Resolution Using Dense Skip Connections ↩

11. Accurate Image Super-Resolution Using Very Deep Convolutional Networks ↩