Neural Sparse Representation for Image Restoration

Preprint [arXiv]

Usage

Clone our Image Restoration Toolkit and install dependencies.

git clone --recurse-submodules https://github.com/ychfan/pt_ir.git

Performance

Image super-resolution

Small models

Networks	Parameters	DIV2K (val)	Set5	B100	Urban100	Manga109	Pre-trained	Eval cmd	Train cmd
EDSR+NSR x2	4,877,812	34.87	38.11	32.22	32.58	38.95	Download	details python trainer.py --eval_only --dataset div2k --eval_datasets div2k set5 bsds100 urban100 manga109 --scale 2 --model wdsr_nsr --num_blocks 16 --num_residual_units 64 --width_multiplier 1 --num_sparsity_groups 4 --job_dir X --ckpt ./logs/edsr_s_nsr_x2/epoch_30.pth	details python trainer.py --dataset div2k --eval_datasets div2k set5 bsds100 urban100 manga109 --scale 2 --model wdsr_nsr --num_blocks 16 --num_residual_units 64 --width_multiplier 1 --num_sparsity_groups 4 --job_dir ./logs/edsr_s_nsr_x2
EDSR+NSR x3	4,887,637	31.10	34.50	29.14	28.45	33.76	Download	details python trainer.py --eval_only --dataset div2k --eval_datasets div2k set5 bsds100 urban100 manga109 --scale 3 --model wdsr_nsr --num_blocks 16 --num_residual_units 64 --width_multiplier 1 --num_sparsity_groups 4 --job_dir X --ckpt ./logs/edsr_s_nsr_x3/epoch_30.pth	details python trainer.py --dataset div2k --eval_datasets div2k set5 bsds100 urban100 manga109 --scale 3 --model wdsr_nsr --num_blocks 16 --num_residual_units 64 --width_multiplier 1 --num_sparsity_groups 4 --job_dir ./logs/edsr_s_nsr_x3
EDSR+NSR x4	4,901,392	29.10	32.26	27.61	26.18	30.62	Download	details python trainer.py --eval_only --dataset div2k --eval_datasets div2k set5 bsds100 urban100 manga109 --scale 4 --model wdsr_nsr --num_blocks 16 --num_residual_units 64 --width_multiplier 1 --num_sparsity_groups 4 --job_dir X --ckpt ./logs/edsr_s_nsr_x4/epoch_30.pth	details python trainer.py --dataset div2k --eval_datasets div2k set5 bsds100 urban100 manga109 --scale 4 --model wdsr_nsr --num_blocks 16 --num_residual_units 64 --width_multiplier 1 --num_sparsity_groups 4 --job_dir ./logs/edsr_s_nsr_x4