SwinIR: Image Restoration Using Swin Transformer

This repository is the official PyTorch implementation of SwinIR: Image Restoration Using Shifted Window Transformer (arxiv, supp). SwinIR achieves state-of-the-art performance in

bicubic/lighweight/real-world image SR
grayscale/color image denoising
JPEG compression artifact reduction

🚀 🚀 🚀 News:

Sep. 30, 2021: We release the SwinIR-Large model for real-world image SR.
Sep. 23, 2021: Integrated to Huggingface Spaces with Gradio. See Gradio Web Demo.
Sep. 07, 2021: We release the SwinIR training code at KAIR🔥🔥
Sep. 07, 2021: We provide an interactive online Colab demo for real-world image SR 🔥🔥 for comparison with the first practical degradation model BSRGAN (ICCV2021) and a recent model RealESRGAN. Try to super-resolve your own images on Colab!

Real-World Image (x4)	BSRGAN, ICCV2021	Real-ESRGAN	SwinIR (ours)	SwinIR-Large (ours)

* Please zoom in for comparison. Image 1 and Image 2 are randomly chosen from the web and from the RealSRSet+5images dataset, respectively.

Sep. 03, 2021: Add more detailed comparison between SwinIR and a representative CNN-based model RCAN (classical image SR, X4)

Method	Training Set	Training time (8GeForceRTX2080Ti batch=32, iter=500k)	Y-PSNR/Y-SSIM on Manga109	Run time (1GeForceRTX2080Ti, on 256x256 LR image)*	#Params	#FLOPs	Testing memory
RCAN	DIV2K	1.6 days	31.22/0.9173	0.180s	15.6M	850.6G	593.1M
SwinIR	DIV2K	1.8 days	31.67/0.9226	0.539s	11.9M	788.6G	986.8M

* We re-test the runtime when the GPU is idle. We refer to the evluation code here.

Aug. 26, 2021: See our recent work on real-world image SR: a pratical degrdation model BSRGAN, ICCV2021
Aug. 26, 2021: See our recent work on generative modelling of image SR and image rescaling: normalizing-flow-based HCFlow, ICCV2021
Aug. 26, 2021: See our recent work on blind SR: spatially variant kernel estimation (MANet, ICCV2021) and unsupervised kernel estimation (FKP, CVPR2021)

Image restoration is a long-standing low-level vision problem that aims to restore high-quality images from low-quality images (e.g., downscaled, noisy and compressed images). While state-of-the-art image restoration methods are based on convolutional neural networks, few attempts have been made with Transformers which show impressive performance on high-level vision tasks. In this paper, we propose a strong baseline model SwinIR for image restoration based on the Swin Transformer. SwinIR consists of three parts: shallow feature extraction, deep feature extraction and high-quality image reconstruction. In particular, the deep feature extraction module is composed of several residual Swin Transformer blocks (RSTB), each of which has several Swin Transformer layers together with a residual connection. We conduct experiments on three representative tasks: image super-resolution (including classical, lightweight and real-world image super-resolution), image denoising (including grayscale and color image denoising) and JPEG compression artifact reduction. Experimental results demonstrate that SwinIR outperforms state-of-the-art methods on different tasks by up to 0.14~0.45dB, while the total number of parameters can be reduced by up to 67%.

Training
Testing
Results
Citation
License and Acknowledgement

Training

Used training and testing sets can be downloaded as follows:

Task	Training Set	Testing Set
classical/lightweight image SR	DIV2K (800 training images) or DIV2K +Flickr2K (2650 images)	Set5 + Set14 + BSD100 + Urban100 + Manga109 download all
real-world image SR	SwinIR-M (middle size): DIV2K (800 training images) +Flickr2K (2650 images) + OST (alternative link, 10324 images for sky,water,grass,mountain,building,plant,animal) SwinIR-L (large size): DIV2K + Flickr2K + OST + WED(4744 images) + FFHQ (first 2000 images, face) + Manga109 (manga) + SCUT-CTW1500 (first 100 training images, texts) *We use the pionnerring practical degradation model from BSRGAN, ICCV2021	RealSRSet+5images
color/grayscale image denoising	DIV2K (800 training images) + Flickr2K (2650 images) + BSD500 (400 training&testing images) + WED(4744 images)	grayscale: Set12 + BSD68 + Urban100 color: CBSD68 + Kodak24 + McMaster + Urban100 download all
JPEG compression artifact reduction	DIV2K (800 training images) + Flickr2K (2650 images) + BSD500 (400 training&testing images) + WED(4744 images)	grayscale: Classic5 +LIVE1 download all

The training code is at KAIR.

Testing (without preparing datasets)

For your convience, we provide some example datasets (~20Mb) in /testsets. If you just want codes, downloading models/network_swinir.py, utils/util_calculate_psnr_ssim.py and main_test_swinir.py is enough. Following commands will download pretrained models automatically and put them in model_zoo/swinir. All visual results of SwinIR can be downloaded here.

We also provide an online Colab demo for real-world image SR for comparison with the first practical degradation model BSRGAN (ICCV2021) GitHub Stars and a recent model RealESRGAN. Try to test your own images on Colab!

# 001 Classical Image Super-Resolution (middle size)
# Note that --training_patch_size is just used to differentiate two different settings in Table 2 of the paper. Images are NOT tested patch by patch.
# (setting1: when model is trained on DIV2K and with training_patch_size=48)
python main_test_swinir.py --task classical_sr --scale 2 --training_patch_size 48 --model_path model_zoo/swinir/001_classicalSR_DIV2K_s48w8_SwinIR-M_x2.pth --folder_lq testsets/Set5/LR_bicubic/X2 --folder_gt testsets/Set5/HR
python main_test_swinir.py --task classical_sr --scale 3 --training_patch_size 48 --model_path model_zoo/swinir/001_classicalSR_DIV2K_s48w8_SwinIR-M_x3.pth --folder_lq testsets/Set5/LR_bicubic/X3 --folder_gt testsets/Set5/HR
python main_test_swinir.py --task classical_sr --scale 4 --training_patch_size 48 --model_path model_zoo/swinir/001_classicalSR_DIV2K_s48w8_SwinIR-M_x4.pth --folder_lq testsets/Set5/LR_bicubic/X4 --folder_gt testsets/Set5/HR
python main_test_swinir.py --task classical_sr --scale 8 --training_patch_size 48 --model_path model_zoo/swinir/001_classicalSR_DIV2K_s48w8_SwinIR-M_x8.pth --folder_lq testsets/Set5/LR_bicubic/X8 --folder_gt testsets/Set5/HR

# (setting2: when model is trained on DIV2K+Flickr2K and with training_patch_size=64)
python main_test_swinir.py --task classical_sr --scale 2 --training_patch_size 64 --model_path model_zoo/swinir/001_classicalSR_DF2K_s64w8_SwinIR-M_x2.pth --folder_lq testsets/Set5/LR_bicubic/X2 --folder_gt testsets/Set5/HR
python main_test_swinir.py --task classical_sr --scale 3 --training_patch_size 64 --model_path model_zoo/swinir/001_classicalSR_DF2K_s64w8_SwinIR-M_x3.pth --folder_lq testsets/Set5/LR_bicubic/X3 --folder_gt testsets/Set5/HR
python main_test_swinir.py --task classical_sr --scale 4 --training_patch_size 64 --model_path model_zoo/swinir/001_classicalSR_DF2K_s64w8_SwinIR-M_x4.pth --folder_lq testsets/Set5/LR_bicubic/X4 --folder_gt testsets/Set5/HR
python main_test_swinir.py --task classical_sr --scale 8 --training_patch_size 64 --model_path model_zoo/swinir/001_classicalSR_DF2K_s64w8_SwinIR-M_x8.pth --folder_lq testsets/Set5/LR_bicubic/X8 --folder_gt testsets/Set5/HR


# 002 Lightweight Image Super-Resolution (small size)
python main_test_swinir.py --task lightweight_sr --scale 2 --model_path model_zoo/swinir/002_lightweightSR_DIV2K_s64w8_SwinIR-S_x2.pth --folder_lq testsets/Set5/LR_bicubic/X2 --folder_gt testsets/Set5/HR
python main_test_swinir.py --task lightweight_sr --scale 3 --model_path model_zoo/swinir/002_lightweightSR_DIV2K_s64w8_SwinIR-S_x3.pth --folder_lq testsets/Set5/LR_bicubic/X3 --folder_gt testsets/Set5/HR
python main_test_swinir.py --task lightweight_sr --scale 4 --model_path model_zoo/swinir/002_lightweightSR_DIV2K_s64w8_SwinIR-S_x4.pth --folder_lq testsets/Set5/LR_bicubic/X4 --folder_gt testsets/Set5/HR


# 003 Real-World Image Super-Resolution (use --tile 400 if you run out-of-memory)
# (middle size)
python main_test_swinir.py --task real_sr --scale 4 --model_path model_zoo/swinir/003_realSR_BSRGAN_DFO_s64w8_SwinIR-M_x4_GAN.pth --folder_lq testsets/RealSRSet+5images --tile

# (larger size + trained on more datasets)
python main_test_swinir.py --task real_sr --scale 4 --large_model --model_path model_zoo/swinir/003_realSR_BSRGAN_DFOWMFC_s64w8_SwinIR-L_x4_GAN.pth --folder_lq testsets/RealSRSet+5images


# 004 Grayscale Image Deoising (middle size)
python main_test_swinir.py --task gray_dn --noise 15 --model_path model_zoo/swinir/004_grayDN_DFWB_s128w8_SwinIR-M_noise15.pth --folder_gt testsets/Set12
python main_test_swinir.py --task gray_dn --noise 25 --model_path model_zoo/swinir/004_grayDN_DFWB_s128w8_SwinIR-M_noise25.pth --folder_gt testsets/Set12
python main_test_swinir.py --task gray_dn --noise 50 --model_path model_zoo/swinir/004_grayDN_DFWB_s128w8_SwinIR-M_noise50.pth --folder_gt testsets/Set12


# 005 Color Image Deoising (middle size)
python main_test_swinir.py --task color_dn --noise 15 --model_path model_zoo/swinir/005_colorDN_DFWB_s128w8_SwinIR-M_noise15.pth --folder_gt testsets/McMaster
python main_test_swinir.py --task color_dn --noise 25 --model_path model_zoo/swinir/005_colorDN_DFWB_s128w8_SwinIR-M_noise25.pth --folder_gt testsets/McMaster
python main_test_swinir.py --task color_dn --noise 50 --model_path model_zoo/swinir/005_colorDN_DFWB_s128w8_SwinIR-M_noise50.pth --folder_gt testsets/McMaster


# 006 JPEG Compression Artifact Reduction (middle size, using window_size=7 because JPEG encoding uses 8x8 blocks)
python main_test_swinir.py --task jpeg_car --jpeg 10 --model_path model_zoo/swinir/006_CAR_DFWB_s126w7_SwinIR-M_jpeg10.pth --folder_gt testsets/classic5
python main_test_swinir.py --task jpeg_car --jpeg 20 --model_path model_zoo/swinir/006_CAR_DFWB_s126w7_SwinIR-M_jpeg20.pth --folder_gt testsets/classic5
python main_test_swinir.py --task jpeg_car --jpeg 30 --model_path model_zoo/swinir/006_CAR_DFWB_s126w7_SwinIR-M_jpeg30.pth --folder_gt testsets/classic5
python main_test_swinir.py --task jpeg_car --jpeg 40 --model_path model_zoo/swinir/006_CAR_DFWB_s126w7_SwinIR-M_jpeg40.pth --folder_gt testsets/classic5

Results

We achieved state-of-the-art performance on classical/lightweight/real-world image SR, grayscale/color image denoising and JPEG compression artifact reduction. Detailed results can be found in the paper.

Classical Image Super-Resolution (click me)

Lightweight Image Super-Resolution

Real-World Image Super-Resolution

Grayscale Image Deoising

Color Image Deoising

JPEG Compression Artifact Reduction

Citation

@inproceedings{liang2021swinir,
    title={SwinIR: Image Restoration Using Swin Transformer},
    author={Liang, Jingyun and Cao, Jiezhang and Sun, Guolei and Zhang, Kai and Van Gool, Luc and Timofte, Radu},
    booktitle={IEEE International Conference on Computer Vision Workshops},
    year={2021}
}

License and Acknowledgement

This project is released under the Apache 2.0 license. The codes are heavily based on Swin Transformer. We also refer to codes in KAIR and BasicSR. Please also follow their licenses. Thanks for their awesome works.

hanlinwu/SwinIR