ShuffleSR

• Lighter and better image super-resolution algorithm than WDSR

Performance

• Training and validation results on DIV2K dataset

DIV2K Mild X4

name	PSNR	SSIM	Flops	parms	checkpoint
Bicubic	17.6147	0.4268	-	-	-
ShuffleSR	18.7887	0.4857	126.5M	56.5K	ShuffleSR_Mild_X4.pth
ShuffleSR_SE	18.8280	0.5078	128.2M	158.9K	ShuffleSR_SE_Mild_X4.pth
ShuffleSR_SK	18.7890	0.5094	165.0M	152.2K	ShuffleSR_SK_Mild_X4.pth
WDSR	18.7757	0.4975	381.5M	167.7K	WDSR_Mild_X4.pth

DIV2K Unknown X4

name	PSNR	SSIM	Flops	parms	checkpoint
Bicubic	21.7511	0.5561	-	-	-
ShuffleSR	24.8454	0.7025	129.2M	58.4K	ShuffleSR_Unknown_X4.pth
ShuffleSR_SK	26.4377	0.7542	165.0M	152.2K	ShuffleSR_SK_Unknown_X4.pth
ShuffleSR_SE	25.6780	0.7292	128.2M	158.9K	ShuffleSR_SE_Unknown_X4.pth
WDSR	25.2512	0.7152	381.5M	167.7K	WDSR_Unknown_X4.pth

DIV2K Bicubic X4

name	PSNR	SSIM	Flops	parms	checkpoint
Bicubic	26.6952	0.7663	-	-	-
ShuffleSR	28.2402	0.8135	126.5M	56.5K	ShuffleSR_Bicubic_X4.pth
ShuffleSR_SK	28.3443	0.8170	165.0M	152.2K	ShuffleSR_SK_Bicubic_X4.pth
WDSR	28.4477	0.8186	381.5M	167.7K	WDSR_Bicubic_X4.pth
ShuffleSR_M	28.4664	0.8192	342.4M	152.1K	ShuffleSR_M_Bicubic_X4.pth

DIV2K Bicubic X2

name	PSNR	SSIM	Flops	parms	checkpoint
Bicubic	31.0374	0.9015	-	-	-
ShuffleSR	33.7818	0.9358	122.0M	54.8K	ShuffleSR_Bicubic_X2.pth
ShuffleSR_SK	33.8937	0.9368	168.2M	188.0K	ShuffleSR_SK_Bicubic_X2.pth
WDSR	33.9800	0.9373	351.2M	154.5K	WDSR_Bicubic_X2.pth
ShuffleSR_M	34.0213	0.9376	321.1M	143.0K	ShuffleSR_M_Bicubic_X2.pth

Network super parameter

network	features	expand	blocks
ShuffleSR_SE_Unknown_X4	40	2	8
ShuffleSR_SE_Mild_X4	40	2	8
ShuffleSR_SK_Bicubic_X2	48	2	8
ShuffleSR_SK_Bicubic_X4	40	2	8
ShuffleSR_SK_Unknown_X4	40	2	8
ShuffleSR_SK_Mild_X4	40	2	8
ShuffleSR_Bicubic_X2	48	2	8
ShuffleSR_Bicubic_X4	40	2	8
ShuffleSR_Unknown_X4	28	3	12
ShuffleSR_Mild_X4	40	2	8
ShuffleSR_M_Bicubic_X2	68	3	8
ShuffleSR_M_Bicubic_X4	68	3	8
WDSR_Bicubic_X2	32	6	8
WDSR_Bicubic_X4	32	6	8
WDSR_Unknown_X4	32	6	8
WDSR_Mild_X4	32	6	8

Usage

• Git clone

cd /path/to/ShuffleSR

• Install requirement

pip install -r requirement.txt

• Download data set DIV2K

Replace '/path/to/xx' in the code with your path

• Train

help:
python train.py -h

examples:
# train new model
python train.py --network ShuffleSR_SE --features 32 --expand 3 --block 8
python train.py --network ShuffleSR --features 32 --expand 6 --block 8 --lr 0.001
python train.py --network WDSR --features 32 --expand 3 --block 8 --epochs 100

# load model and retrain
python train.py --load_model_path checkpoint/ShuffleSR_Bicubic_X2.pth
python train.py --load_model_path checkpoint/WDSR_Mild_X4.pth

• Test

set model_path and output_path in the output.py

python output.py

ShuffleSR SR result

DIV2K Unknown X4

• 0825.png

• 0826.png

• 0827.png

• 0832.png

DIV2K Mild X4

• 0836.png

• 0844.png

• 0898.png