Super-resolution is a process that increases the resolution of an image, adding additional details. Methods using neural networks give the most accurate results, much better than other interpolation methods. With the right training, it is even possible to make photo-realistic images.
For example, here is a low-resolution image, magnified x4 by a neural network, and a high resolution image of the same object:
In this repository, you will find:
- the popular super-resolution networks, pretrained
- common super-resolution datasets
- a unified training script for all models
The following pretrained models are available. Click on the links for the paper:
Newer and larger models perform better: the most accurate models are EDSR (huge), RCAN and NinaSR-B2. For practical applications, I recommend a smaller model, such as NinaSR-B1.
Expand benchmark results
Set5 results
Network | Parameters (M) | 2x (PSNR/SSIM) | 3x (PSNR/SSIM) | 4x (PSNR/SSIM) |
---|---|---|---|---|
carn | 1.59 | 37.88 / 0.9600 | 34.32 / 0.9265 | 32.14 / 0.8942 |
carn_m | 0.41 | 37.68 / 0.9594 | 34.06 / 0.9247 | 31.88 / 0.8907 |
edsr_baseline | 1.37 | 37.98 / 0.9604 | 34.37 / 0.9270 | 32.09 / 0.8936 |
edsr | 40.7 | 38.19 / 0.9609 | 34.68 / 0.9293 | 32.48 / 0.8985 |
ninasr_b0 | 0.10 | 37.69 / 0.9594 | 33.91 / 0.9229 | 31.65 / 0.8868 |
ninasr_b1 | 1.02 | 38.00 / 0.9604 | 34.42 / 0.9274 | 32.21 / 0.8947 |
ninasr_b2 | 10.0 | 38.22 / 0.9612 | 34.63 / 0.9288 | 32.48 / 0.8976 |
rcan | 15.4 | 38.27 / 0.9614 | 34.76 / 0.9299 | 32.64 / 0.9000 |
rdn | 22.1 | 38.12 / 0.9609 | 33.98 / 0.9234 | 32.35 / 0.8968 |
Set14 results
Network | Parameters (M) | 2x (PSNR/SSIM) | 3x (PSNR/SSIM) | 4x (PSNR/SSIM) |
---|---|---|---|---|
carn | 1.59 | 33.57 / 0.9173 | 30.30 / 0.8412 | 28.61 / 0.7806 |
carn_m | 0.41 | 33.30 / 0.9151 | 30.10 / 0.8374 | 28.42 / 0.7764 |
edsr_baseline | 1.37 | 33.57 / 0.9174 | 30.28 / 0.8414 | 28.58 / 0.7804 |
edsr | 40.7 | 33.95 / 0.9201 | 30.53 / 0.8464 | 28.81 / 0.7872 |
ninasr_b0 | 0.10 | 33.23 / 0.9147 | 30.01 / 0.8352 | 28.26 / 0.7723 |
ninasr_b1 | 1.02 | 33.61 / 0.9176 | 30.37 / 0.8430 | 28.65 / 0.7824 |
ninasr_b2 | 10.0 | 33.99 / 0.9206 | 30.55 / 0.8461 | 28.81 / 0.7865 |
rcan | 15.4 | 34.13 / 0.9216 | 30.63 / 0.8475 | 28.85 / 0.7878 |
rdn | 22.1 | 33.71 / 0.9182 | 30.07 / 0.8373 | 28.72 / 0.7846 |
DIV2K results (validation set)
Network | Parameters (M) | 2x (PSNR/SSIM) | 3x (PSNR/SSIM) | 4x (PSNR/SSIM) | 8x (PSNR/SSIM) |
---|---|---|---|---|---|
carn | 1.59 | 36.08 / 0.9451 | 32.37 / 0.8871 | 30.43 / 0.8366 | N/A |
carn_m | 0.41 | 35.76 / 0.9429 | 32.09 / 0.8827 | 30.18 / 0.8313 | N/A |
edsr_baseline | 1.37 | 36.13 / 0.9455 | 32.41 / 0.8878 | 30.43 / 0.8370 | N/A |
edsr | 40.7 | 36.56 / 0.9485 | 32.75 / 0.8933 | 30.73 / 0.8445 | N/A |
ninasr_b0 | 0.10 | 35.72 / 0.9424 | 32.01 / 0.8811 | 30.08 / 0.8289 | 26.58 / 0.7076 |
ninasr_b1 | 1.02 | 36.23 / 0.9463 | 32.49 / 0.8891 | 30.53 / 0.8394 | 26.92 / 0.7195 |
ninasr_b2 | 10.0 | 36.54 / 0.9484 | 32.74 / 0.8927 | 30.74 / 0.8441 | 27.07 / 0.7247 |
rcan | 15.4 | 36.61 / 0.9489 | 32.78 / 0.8935 | 30.73 / 0.8447 | 27.17 / 0.7292 |
rdn | 22.1 | 36.32 / 0.9468 | 32.04 / 0.8822 | 30.61 / 0.8414 | N/A |
B100 results
Network | Parameters (M) | 2x (PSNR/SSIM) | 3x (PSNR/SSIM) | 4x (PSNR/SSIM) |
---|---|---|---|---|
carn | 1.59 | 32.12 / 0.8986 | 29.07 / 0.8042 | 27.58 / 0.7355 |
carn_m | 0.41 | 31.97 / 0.8971 | 28.94 / 0.8010 | 27.45 / 0.7312 |
edsr_baseline | 1.37 | 32.15 / 0.8993 | 29.08 / 0.8051 | 27.56 / 0.7354 |
edsr | 40.7 | 32.35 / 0.9019 | 29.26 / 0.8096 | 27.72 / 0.7419 |
ninasr_b0 | 0.10 | 31.94 / 0.8969 | 28.87 / 0.7996 | 27.35 / 0.7285 |
ninasr_b1 | 1.02 | 32.19 / 0.8999 | 29.11 / 0.8056 | 27.60 / 0.7369 |
ninasr_b2 | 10.0 | 32.34 / 0.9018 | 29.25 / 0.8090 | 27.71 / 0.7411 |
rcan | 15.4 | 32.39 / 0.9024 | 29.30 / 0.8106 | 27.74 / 0.7429 |
rdn | 22.1 | 32.25 / 0.9006 | 28.90 / 0.8004 | 27.66 / 0.7388 |
Urban100 results
Network | Parameters (M) | 2x (PSNR/SSIM) | 3x (PSNR/SSIM) | 4x (PSNR/SSIM) |
---|---|---|---|---|
carn | 1.59 | 31.95 / 0.9263 | 28.07 / 0.849 | 26.07 / 0.78349 |
carn_m | 0.41 | 31.30 / 0.9200 | 27.57 / 0.839 | 25.64 / 0.76961 |
edsr_baseline | 1.37 | 31.98 / 0.9271 | 28.15 / 0.852 | 26.03 / 0.78424 |
edsr | 40.7 | 32.97 / 0.9358 | 28.81 / 0.865 | 26.65 / 0.80328 |
ninasr_b0 | 0.10 | 31.21 / 0.9190 | 27.37 / 0.834 | 25.40 / 0.76207 |
ninasr_b1 | 1.02 | 32.18 / 0.9288 | 28.23 / 0.854 | 26.11 / 0.78772 |
ninasr_b2 | 10.0 | 32.92 / 0.9356 | 28.69 / 0.863 | 26.55 / 0.80087 |
rcan | 15.4 | 33.19 / 0.9372 | 29.01 / 0.868 | 26.75 / 0.80624 |
rdn | 22.1 | 32.41 / 0.9310 | 27.49 / 0.838 | 26.36 / 0.79460 |
All models are defined in torchsr.models
. Other useful tools to augment your models, such as self-ensemble methods and tiling, are present in torchsr.models.utils
.
The following datasets are available. Click on the links for the project page:
All datasets are defined in torchsr.datasets
. They return a list of images, with the high-resolution image followed by downscaled or degraded versions.
Data augmentation methods are provided in torchsr.transforms
.
Datasets are downloaded automatically when using the download=True
flag, or by running the corresponding script i.e. ./scripts/download_div2k.sh
.
from torchsr.datasets import Div2K
from torchsr.models import ninasr_b0
from torchvision.transforms.functional import to_pil_image, to_tensor
# Div2K dataset
dataset = Div2K(root="./data", scale=2, download=False)
# Get the first image in the dataset (High-Res and Low-Res)
hr, lr = dataset[0]
# Download a pretrained NinaSR model
model = ninasr_b0(scale=2, pretrained=True)
# Run the Super-Resolution model
lr_t = to_tensor(lr).unsqueeze(0)
sr_t = model(lr_t)
sr = to_pil_image(sr_t.squeeze(0))
sr.show()
Expand more examples
from torchsr.datasets import Div2K
from torchsr.models import edsr, rcan
from torchsr.models.utils import ChoppedModel, SelfEnsembleModel
from torchsr.transforms import ColorJitter, Compose, RandomCrop
# Div2K dataset, cropped to 256px, width color jitter
dataset = Div2K(
root="./data", scale=2, download=False,
transform=Compose([
RandomCrop(256, scales=[1, 2]),
ColorJitter(brightness=0.2)
]))
# Pretrained RCAN model, with tiling for large images
model = ChoppedModel(
rcan(scale=2, pretrained=True), scale=2,
chop_size=400, chop_overlap=10)
# Pretrained EDSR model, with self-ensemble method for higher quality
model = SelfEnsembleModel(edsr(scale=2, pretrained=True))
A script is available to train the models from scratch, evaluate them, and much more. It is not part of the pip package, and requires additional dependencies.
pip install piq tqdm tensorboard # Additional dependencies
python -m torchsr.train -h
python -m torchsr.train --arch edsr_baseline --scale 2 --download-pretrained --images test/butterfly.png --destination results/
python -m torchsr.train --arch edsr_baseline --scale 2 --download-pretrained --validation-only
python -m torchsr.train --arch edsr_baseline --scale 2 --epochs 300 --loss l1 --dataset-train div2k_bicubic
Thanks to the people behind torchvision and EDSR, whose work inspired this repository. Some of the models available here come from EDSR-PyTorch and CARN-PyTorch.
To cite this work, please use:
@misc{torchsr,
author = {Gabriel Gouvine},
title = {Super-Resolution Networks for Pytorch},
year = {2021},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/Coloquinte/torchSR}},
doi = {10.5281/zenodo.4868308}
}