SwinIR: Image Restoration Using Swin Transformer

Jingyun LiangJiezhang CaoGuolei SunKai ZhangLuc Van GoolRadu Timofte

article2021ICCV5,134 citations

Proposes SwinIR, a Swin Transformer-based architecture that outperforms state-of-the-art convolutional networks across super-resolution, denoising, and JPEG artifact removal while using up to 67% fewer parameters.

Listen

Digital image restoration—including super-resolution, denoising, and compression artifact removal—is essential for recovering high-quality visual data from degraded inputs across consumer, industrial, and medical applications. While convolutional neural networks have long served as the standard approach, they struggle to model long-range contextual relationships and process all image regions using fixed, content-independent filters. Early attempts to apply self-attention transformer models to image restoration suffered from heavy computational burdens, artificial border artifacts caused by patch-by-patch processing, and an extreme reliance on massive datasets and hundreds of millions of parameters.

The article evaluates whether an adapted vision transformer architecture, termed SwinIR, can overcome these limitations to serve as a versatile, parameter-efficient baseline for image restoration. The authors set out to demonstrate that shifting-window self-attention combined with residual convolutional layers can outperform existing convolutional and transformer-based methods across diverse restoration tasks.

To test this, the authors designed a three-part modular pipeline consisting of shallow feature extraction, deep feature extraction using residual transformer blocks, and task-specific high-quality image reconstruction. The framework was evaluated across three core problem domains: image super-resolution (classical, lightweight, and real-world), image denoising (grayscale and color), and JPEG compression artifact reduction. Performance was measured using standard signal quality metrics (peak signal-to-noise ratio and structural similarity) across multiple public benchmark datasets, including Set5, Set14, Urban100, Manga109, and LIVE1, against leading convolutional and transformer alternatives.

The experimental findings show that the proposed architecture establishes a new performance baseline. Across classical super-resolution tasks, SwinIR outperformed existing state-of-the-art methods by up to 0.45 dB, restoring sharper edges and rich high-frequency textures where alternative methods produced blurred or distorted patterns. On lightweight super-resolution, it achieved superior restoration accuracy while maintaining comparable computational operations and parameter counts. In image denoising and JPEG artifact reduction, the model exceeded leading models like DRUNet and DnCNN, delivering up to a 0.3 dB improvement while requiring up to 67% fewer parameters (11.5–12.0 million compared to 32.7 million). Furthermore, training analysis demonstrated that the model converged faster and achieved higher accuracy than deep convolutional networks even when trained on small datasets of only a few hundred images, refuting the assumption that vision transformers require millions of pre-training images to be effective.

These findings indicate that organizations deploying automated computer vision and image processing systems can achieve noticeably higher reconstruction quality with substantially smaller model footprints. The reduced parameter footprint lowers the computational and hardware costs required for model deployment in cloud and edge environments without sacrificing fidelity. Moreover, the architecture's ability to generalize across six different restoration settings indicates that teams can maintain a unified technical framework rather than building fragmented pipelines for individual degradation problems.

Based on these results, engineering and technical teams should consider adopting this shifted-window transformer framework as a primary architecture for image enhancement workflows. For deployment in resource-constrained environments, the lightweight variant provides an immediate efficiency trade-off. Before broad commercial deployment, organizations should conduct domain-specific pilot testing on real-world imagery, adjusting the restoration modules and training data to mirror their specific corruption profiles. Future research and development should focus on expanding the architecture to handle other degradation challenges, such as image deblurring and deraining.

Confidence in the findings is high for standard image benchmarks, as the model was rigorously tested across multiple tasks, noise levels, and scaling factors. However, users should note that the base model was primarily evaluated against synthetic degradation pipelines and standard benchmark sets; performance on complex real-world corruptions may require larger training sets and task-specific loss configurations to prevent unmodeled visual artifacts.

ETH Zurich KU Leuven color images compressed images convolutional neural networks Swin Transformer image restoration image superresolution

arXiv: 2108.10257 JingyunLiang/SwinIR

Paper: Swin Transformer: Hierarchical Vision Transformer using Shifted Windows, Ze Liu et al. (2021). Reading the Swin Transformer paper first is essential because SwinIR directly builds its deep feature extraction blocks on top of the Swin Transformer architecture.
Paper: Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising, Kai Zhang et al. (2016). Understanding this foundational work on residual learning for image denoising helps clarify the architectural strategies SwinIR adapts for restoration tasks.
Paper: Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network, Christian Ledig et al. (2017). Familiarizing yourself with this benchmark restoration study provides necessary context for evaluating the image super-resolution tasks addressed by SwinIR.

Paper: Restormer: Efficient Transformer for High-Resolution Image Restoration, Syed Waqas Zamir et al. (2022). Restormer naturally follows SwinIR by extending Transformer-based image restoration to handle higher-resolution images and diverse degradation types even more efficiently.
Paper: Swin Transformer V2: Scaling Up Capacity and Resolution, Ze Liu et al. (2022). Swin Transformer V2 serves as a direct follow-up that addresses the scaling limits of the original Swin architecture used in SwinIR.

Cover for SwinIR: Image Restoration Using Swin Transformer

Abstract

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 $\textbf{up to 0.14$ \sim $0.45dB}$ , while the total number of parameters can be reduced by $\textbf{up to 67%}$ .

Topics

color images compressed images convolutional neural networks encoder-decoder feature extraction image denoising jpeg jpeg compression reconstruction restoration methods

Keywords

Swin Transformer image restoration image superresolution image denoising JPEG compression artifact reduction blind image super-resolution Vision Transformer

1 Introduction
2 Related Work
2.1 Image Restoration
2.2 Vision Transformer
3 Method
3.1 Network Architecture
3.2 Residual Swin Transformer Block
4 Experiments
4.1 Experimental Setup
4.2 Ablation Study and Discussion
4.3 Results on Image SR
4.4 Results on JPEG Compression Artifact Reduction
4.5 Results on Image Denoising
5 Conclusion
References

Knowls

Knowl 1 — SwinIR Overall Architecture for Image Restoration
model/method
SwinIR is an image restoration network based on the Swin Transformer. It comprises three sequential modules: shallow feature extraction, deep feature extraction, and high-quality (HQ) image reconstruction.
1. Shallow Feature Extraction: Given a low-quality (LQ) image $I_{\text{LQ}} \in \mathbb{R}^{H \times W \times C_{\text{in}}}$ (with spatial dimensions $H \times W$ and channel count $C_{\text{in}}$ ), a $3 \times 3$ convolutional layer $H_{\text{SF}}(\cdot)$ extracts shallow features $F_0 \in \mathbb{R}^{H \times W \times C}$ :
$F_0 = H_{\text{SF}}(I_{\text{LQ}})$
1. Deep Feature Extraction: Deep features $F_{\text{DF}} \in \mathbb{R}^{H \times W \times C}$ are extracted using $K$ residual Swin Transformer blocks (RSTB) followed by a final $3 \times 3$ convolutional layer $H_{\text{CONV}}(\cdot)$ :
$\begin{aligned} F_i &= H_{\text{RSTB}_i}(F_{i-1}), \quad i = 1, 2, \dots, K \\ F_{\text{DF}} &= H_{\text{CONV}}(F_K) \end{aligned}$
1. High-Quality Image Reconstruction: For super-resolution (SR), shallow and deep features are aggregated via a long skip connection and processed through an upsampling reconstruction module $H_{\text{REC}}(\cdot)$ (implemented via sub-pixel convolution layers):
$I_{\text{RHQ}} = H_{\text{REC}}(F_0 + F_{\text{DF}})$

For non-upsampling tasks such as image denoising and JPEG compression artifact reduction, a single $3 \times 3$ convolution layer is used in conjunction with global residual learning to predict the residual between the input and ground-truth images:

$I_{\text{RHQ}} = H_{\text{SwinIR}}(I_{\text{LQ}}) + I_{\text{LQ}}$
Knowl 2 — Residual Swin Transformer Block (RSTB)
model/method
The Residual Swin Transformer Block (RSTB) is the core building block of deep feature extraction in SwinIR. For the $i$ -th RSTB receiving input feature $F_{i,0} \in \mathbb{R}^{H \times W \times C}$ , feature transformation occurs in two stages:
1. Sequential Swin Transformer Layers: A series of $L$ Swin Transformer layers (STL) extract intermediate representations:
$F_{i,j} = H_{\text{Swin}_{i,j}}(F_{i,j-1}), \quad j = 1, 2, \dots, L$
1. Convolution and Residual Shortcut: The output of the final STL passes through a $3 \times 3$ convolution layer $H_{\text{CONV}_i}(\cdot)$ , and the original block input $F_{i,0}$ is added via an identity connection:
$F_{i,\text{out}} = H_{\text{CONV}_i}(F_{i,L}) + F_{i,0}$

The $3 \times 3$ convolution layer introduces spatial translation equivariance and the inductive bias of local operations to the Transformer representation, while the residual connection provides a shortcut that facilitates multi-level feature aggregation.
Knowl 3 — Swin Transformer Layer (STL) Formulation in SwinIR
model/method
Each Swin Transformer layer (STL) processes an input feature map $X \in \mathbb{R}^{H \times W \times C}$ by partitioning it into non-overlapping local windows of size $M \times M$ , yielding $\frac{HW}{M^2}$ local patches of dimension $M^2 \times C$ . For each local window feature matrix $X \in \mathbb{R}^{M^2 \times C}$ , query ( $Q$ ), key ( $K$ ), and value ( $V$ ) projections are computed using shared linear transformation matrices $P_Q, P_K, P_V \in \mathbb{R}^{C \times d}$ :

$Q = X P_Q, \quad K = X P_K, \quad V = X P_V$

Local window self-attention is computed with a learnable relative positional encoding matrix $B \in \mathbb{R}^{M^2 \times M^2}$ :

$\text{Attention}(Q, K, V) = \text{SoftMax}\left(\frac{Q K^T}{\sqrt{d}} + B\right)V$

Multi-head self-attention (MSA) runs $h$ attention heads in parallel and concatenates their outputs. The full layer transformation, including LayerNorm (LN) and a multi-layer perceptron (MLP) containing two fully connected layers with GELU non-linearities, is defined as:

$\begin{aligned} X &= \text{MSA}(\text{LN}(X)) + X \\ X &= \text{MLP}(\text{LN}(X)) + X \end{aligned}$

To allow information exchange across different local windows, consecutive STLs alternate between standard local window partitioning and shifted window partitioning (shifting features by $(\lfloor M/2 \rfloor, \lfloor M/2 \rfloor)$ pixels before window division).
Knowl 4 — Loss Functions for Restoration Objectives in SwinIR
equation
SwinIR employs task-specific loss functions between the restored image $I_{\text{RHQ}}$ and ground truth $I_{\text{HQ}}$ :
1. Classical and Lightweight Image Super-Resolution: The network is optimized using the pixel-level $L_1$ loss:
$\mathcal{L} = \|I_{\text{RHQ}} - I_{\text{HQ}}\|_1$
1. Image Denoising and JPEG Compression Artifact Reduction: The network is optimized using the Charbonnier loss:
$\mathcal{L} = \sqrt{\|I_{\text{RHQ}} - I_{\text{HQ}}\|^2 + \epsilon^2}$

where $\epsilon$ is a constant empirically set to $10^{-3}$ .
1. Real-World Super-Resolution: The model is trained using a weighted combination of pixel loss ( $L_1$ ), GAN loss, and perceptual loss to produce visually sharp textures.

Knowl 5 — Quantitative Evaluation on Classical Image Super-Resolution

data/table

SwinIR (configured with $K=6$ RSTBs, $L=6$ STLs per block, channel dimension $C=180$ , attention heads $h=6$ , window size $M=8$ , total parameters $\approx 11.8\text{M}$ ) was evaluated across standard classical image SR benchmarks against representative CNN and Transformer models.

Method	Scale	Training Dataset	Set5	Set14	BSD100	Urban100	Manga109
			PSNR / SSIM	PSNR / SSIM	PSNR / SSIM	PSNR / SSIM	PSNR / SSIM
RCAN	$\times 2$	DIV2K	38.27 / 0.9614	34.12 / 0.9216	32.41 / 0.9027	33.34 / 0.9384	39.44 / 0.9786
SAN	$\times 2$	DIV2K	38.31 / 0.9620	34.07 / 0.9213	32.42 / 0.9028	33.10 / 0.9370	39.32 / 0.9792
NLSA	$\times 2$	DIV2K	38.34 / 0.9618	34.08 / 0.9231	32.43 / 0.9027	33.42 / 0.9394	39.59 / 0.9789
SwinIR	$\times 2$	DIV2K	38.35 / 0.9620	34.14 / 0.9227	32.44 / 0.9030	33.40 / 0.9393	39.60 / 0.9792
SwinIR+	$\times 2$	DIV2K	38.38 / 0.9621	34.24 / 0.9233	32.47 / 0.9032	33.51 / 0.9401	39.70 / 0.9794
IPT	$\times 2$	ImageNet	38.37 / -	34.43 / -	32.48 / -	33.76 / -	- / -
SwinIR	$\times 2$	DIV2K+Flickr2K	38.42 / 0.9623	34.46 / 0.9250	32.53 / 0.9041	33.81 / 0.9427	39.92 / 0.9797
SwinIR+	$\times 2$	DIV2K+Flickr2K	38.46 / 0.9624	34.61 / 0.9260	32.55 / 0.9043	33.95 / 0.9433	40.02 / 0.9800
RCAN	$\times 4$	DIV2K	32.63 / 0.9002	28.87 / 0.7889	27.77 / 0.7436	26.82 / 0.8087	31.22 / 0.9173
SAN	$\times 4$	DIV2K	32.64 / 0.9003	28.92 / 0.7888	27.78 / 0.7436	26.79 / 0.8068	31.18 / 0.9169
HAN	$\times 4$	DIV2K	32.64 / 0.9002	28.90 / 0.7890	27.80 / 0.7442	26.85 / 0.8094	31.42 / 0.9177
NLSA	$\times 4$	DIV2K	32.59 / 0.9000	28.87 / 0.7891	27.78 / 0.7444	26.96 / 0.8109	31.27 / 0.9184
SwinIR	$\times 4$	DIV2K	32.72 / 0.9021	28.94 / 0.7914	27.83 / 0.7459	27.07 / 0.8164	31.67 / 0.9226
SwinIR+	$\times 4$	DIV2K	32.81 / 0.9029	29.02 / 0.7928	27.87 / 0.7466	27.21 / 0.8187	31.88 / 0.9423
IPT	$\times 4$	ImageNet	32.64 / -	29.01 / -	27.82 / -	27.26 / -	- / -
RRDB	$\times 4$	DIV2K+Flickr2K	32.73 / 0.9011	28.99 / 0.7917	27.85 / 0.7455	27.03 / 0.8153	31.66 / 0.9196
SwinIR	$\times 4$	DIV2K+Flickr2K	32.92 / 0.9044	29.09 / 0.7950	27.92 / 0.7489	27.45 / 0.8254	32.03 / 0.9260
SwinIR+	$\times 4$	DIV2K+Flickr2K	32.93 / 0.9043	29.15 / 0.7958	27.95 / 0.7494	27.56 / 0.8273	32.22 / 0.9273

When trained on DIV2K+Flickr2K, SwinIR surpasses IPT (which uses 115.5M parameters and pretraining on over 1.1M ImageNet images) across all scale factors while using only 11.8M parameters (an $\approx 90\%$ parameter reduction compared to IPT).

Knowl 6 — Quantitative Evaluation on Lightweight Image Super-Resolution

data/table

For lightweight single-image super-resolution, SwinIR was configured with $K=4$ RSTBs, $L=6$ STLs per block, and channel dimension $C=60$ . Multi-Adds were evaluated on a $1280 \times 720$ output HQ image.

Method	Scale	Params	Mult-Adds	Set5	Set14	BSD100	Urban100	Manga109
				PSNR / SSIM	PSNR / SSIM	PSNR / SSIM	PSNR / SSIM	PSNR / SSIM
CARN	$\times 2$	1,592K	222.8G	37.76 / 0.9590	33.52 / 0.9166	32.09 / 0.8978	31.92 / 0.9256	38.36 / 0.9765
IMDN	$\times 2$	694K	158.8G	38.00 / 0.9605	33.63 / 0.9177	32.19 / 0.8996	32.17 / 0.9283	38.88 / 0.9774
LAPAR-A	$\times 2$	548K	171.0G	38.01 / 0.9605	33.62 / 0.9183	32.19 / 0.8999	32.10 / 0.9283	38.67 / 0.9772
LatticeNet	$\times 2$	756K	169.5G	38.15 / 0.9610	33.78 / 0.9193	32.25 / 0.9005	32.43 / 0.9302	- / -
SwinIR	$\times 2$	878K	195.6G	38.14 / 0.9611	33.86 / 0.9206	32.31 / 0.9012	32.76 / 0.9340	39.12 / 0.9783
CARN	$\times 4$	1,592K	90.9G	32.13 / 0.8937	28.60 / 0.7806	27.58 / 0.7349	26.07 / 0.7837	30.47 / 0.9084
IMDN	$\times 4$	715K	40.9G	32.21 / 0.8948	28.58 / 0.7811	27.56 / 0.7353	26.04 / 0.7838	30.45 / 0.9075
LAPAR-A	$\times 4$	659K	94.0G	32.15 / 0.8944	28.61 / 0.7818	27.61 / 0.7366	26.14 / 0.7871	30.42 / 0.9074
LatticeNet	$\times 4$	777K	43.6G	32.30 / 0.8962	28.68 / 0.7830	27.62 / 0.7367	26.25 / 0.7873	- / -
SwinIR	$\times 4$	897K	49.6G	32.44 / 0.8976	28.77 / 0.7858	27.69 / 0.7406	26.47 / 0.7980	30.92 / 0.9151

SwinIR outperforms previous lightweight models (e.g., CARN, IMDN, LatticeNet) across all benchmarks, gaining up to $0.53\text{ dB}$ on Manga109 for $\times 4$ SR while maintaining comparable parameter size and operation counts.

Knowl 7 — Quantitative Evaluation on Grayscale and Color Image Denoising

data/table

SwinIR was evaluated for additive white Gaussian noise removal at noise levels $\sigma \in \{15, 25, 50\}$ . SwinIR utilizes $\approx 12.0\text{M}$ parameters for denoising compared to $32.7\text{M}$ in DRUNet (a $63\%$ reduction).

Task / Dataset	$\sigma$	DnCNN	FFDNet	NLRN	RNAN	DRUNet	SwinIR
Grayscale Set12	15	32.86	32.75	33.16	-	33.25	33.36
	25	30.44	30.43	30.80	-	30.94	31.01
	50	27.18	27.32	27.64	27.70	27.90	27.91
Grayscale Urban100	15	32.64	32.40	33.45	-	33.44	33.70
	25	29.95	29.90	30.94	-	31.11	31.30
	50	26.26	26.50	27.49	27.65	27.96	27.98
Color CBSD68	15	33.90	33.87	-	-	34.30	34.42
	25	31.24	31.21	-	-	31.69	31.78
	50	27.95	27.96	-	28.27	28.51	28.56
Color Urban100	15	32.98	33.83	-	-	34.81	35.13
	25	30.81	31.40	-	-	32.60	32.90
	50	27.59	28.05	-	29.08	29.61	29.82

SwinIR achieves superior PSNR performance across all noise levels, surpassing DRUNet by up to $0.32\text{ dB}$ on Urban100 for color denoising at $\sigma = 15$ .

Knowl 8 — Quantitative Evaluation on JPEG Compression Artifact Reduction

data/table

SwinIR was evaluated for JPEG deblocking across compression quality factors $q \in \{10, 20, 30, 40\}$ on Classic5 and LIVE1 datasets. For this task, SwinIR adopts a window size of $M=7$ (instead of $M=8$ ) to avoid window resonance with the standard $8 \times 8$ JPEG block grid.

Dataset	$q$	ARCNN	RNAN	RDN	DRUNet	SwinIR
		PSNR / SSIM	PSNR / SSIM	PSNR / SSIM	PSNR / SSIM	PSNR / SSIM
Classic5	10	29.03 / 0.7929	29.96 / 0.8178	30.00 / 0.8188	30.16 / 0.8234	30.27 / 0.8249
	20	31.15 / 0.8517	32.11 / 0.8693	32.15 / 0.8699	32.39 / 0.8734	32.52 / 0.8748
	30	32.51 / 0.8806	33.38 / 0.8924	33.43 / 0.8930	33.59 / 0.8949	33.73 / 0.8961
	40	33.32 / 0.8953	34.27 / 0.9061	34.27 / 0.9061	34.41 / 0.9075	34.52 / 0.9082
LIVE1	10	28.96 / 0.8076	29.63 / 0.8239	29.67 / 0.8247	29.79 / 0.8278	29.86 / 0.8287
	20	31.29 / 0.8733	32.03 / 0.8877	32.07 / 0.8882	32.17 / 0.8899	32.25 / 0.8909
	30	32.67 / 0.9043	33.45 / 0.9149	33.51 / 0.9153	33.59 / 0.9166	33.69 / 0.9174
	40	33.63 / 0.9198	34.47 / 0.9299	34.51 / 0.9302	34.58 / 0.9312	34.67 / 0.9317

SwinIR outperforms DRUNet across all quality levels, achieving average PSNR improvements of at least $0.11\text{ dB}$ on Classic5 and $0.07\text{ dB}$ on LIVE1, while having $11.5\text{M}$ parameters compared to DRUNet's $32.7\text{M}$ .

Knowl 9 — Ablation of RSTB Internal Design and Model Hyperparameters
empirical result
Ablation experiments conducted on Manga109 for classical $\times 2$ super-resolution evaluated the design choices within the Residual Swin Transformer Block (RSTB) and network hyperparameters:
1. Residual Connection and Convolution in RSTB:
  - Without residual connection: $39.42\text{ dB}$
  - With residual connection and $1 \times 1$ convolution: $39.45\text{ dB}$
  - With residual connection and three $3 \times 3$ convolutions: $39.56\text{ dB}$
  - With residual connection and a single $3 \times 3$ convolution: $39.58\text{ dB}$ The identity residual shortcut contributes $+0.16\text{ dB}$ , and the $3 \times 3$ convolution outperforms $1 \times 1$ convolution by extracting local spatial context.
2. Hyperparameter Selection:
  - Channel dimension $C$ : Increasing $C$ improves PSNR, but parameters scale quadratically; $C=180$ was selected as the optimal performance-complexity trade-off.
  - RSTB count $K$ and STL count $L$ : Performance increases and begins saturating around $K=6$ and $L=6$ , which was chosen as the standard middle-size baseline.
Knowl 10 — Convergence Speed and Data Efficiency of SwinIR vs. CNNs
empirical result
In single-image super-resolution comparisons against the representative CNN-based model RCAN on DIV2K and Flickr2K:
1. Data Efficiency: Unlike standard Vision Transformers (e.g., IPT) that require millions of pretraining images to achieve strong performance, SwinIR outperforms RCAN across all training dataset sizes, including small regimes (e.g., $25\%$ of DIV2K, consisting of 200 images).
2. Patch Size Sensitivity: SwinIR outperforms RCAN at all training patch sizes, with the performance gap widening in favor of SwinIR as patch size increases from $32$ to $80$ .
3. Training Convergence: SwinIR converges faster and achieves higher PSNR throughout optimization compared to RCAN, demonstrating that window-based local self-attention avoids the slow convergence typically observed in full-image Vision Transformers.

Coverage note — Qualitative visual comparison figures for super-resolution, denoising, and real-world degradation experiments were omitted in favor of quantitative performance tables and self-contained architectural descriptions.

References

1.Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 126–135, 2017. 4
2.Namhyuk Ahn, Byungkon Kang, and Kyung-Ah Sohn. Fast, accurate, and lightweight super-resolution with cascading residual network. In European Conference on Computer Vision, pages 252–268, 2018. 5, 6
3.Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie line Alberi Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In British Machine Vision Conference, pages 135.1–135.10, 2012. 1, 6
4.Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. arXiv preprint arXiv:2105.05537, 2021. 2
5.Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv preprint arXiv:2106.06847, 2021. 1, 2
6.Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, pages 213–229. Springer, 2020. 1, 2
7.Lukas Cavigelli, Pascal Hager, and Luca Benini. Cas-cnn: A deep convolutional neural network for image compression artifact suppression. In 2017 International Joint Conference on Neural Networks, pages 752–759, 2017. 2
8.Pierre Charbonnier, Laure Blanc-Feraud, Gilles Aubert, and Michel Barlaud. Two deterministic half-quadratic regularization algorithms for computed imaging. In International Conference on Image Processing, volume 2, pages 168–172. IEEE, 1994. 3
9.Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12299–12310, 2021. 1, 2, 5, 6, 8
10.Yunjin Chen and Thomas Pock. Trainable nonlinear reaction diffusion: A flexible framework for fast and effective image restoration. IEEE transactions on pattern analysis and machine intelligence, 39(6):1256–1272, 2016. 2
11.Wenlong Cheng, Mingbo Zhao, Zhiling Ye, and Shuhang Gu. Mfagan: A compression framework for memory-efficient on-device super-resolution gan. arXiv preprint arXiv:2107.12679, 2021. 2
12.Xiangxiang Chu, Bo Zhang, Hailong Ma, Ruijun Xu, and Qingyuan Li. Fast, accurate and lightweight super-resolution with neural architecture search. In International Conference on Pattern Recognition, pages 59–64. IEEE, 2020. 5, 6
13.Jean-Baptiste Cordonnier, Andreas Loukas, and Martin Jaggi. On the relationship between self-attention and convolutional layers. arXiv preprint arXiv:1911.03584, 2019. 2
14.Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering. IEEE Transactions on image processing, 16(8):2080–2095, 2007. 1, 7, 8
15.Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 11065–11074, 2019. 2, 5, 6
16.Xin Deng, Yutong Zhang, Mai Xu, Shuhang Gu, and Yiping Duan. Deep coupled feedback network for joint exposure fusion and image super-resolution. IEEE Transactions on Image Processing, 30:3098–3112, 2021. 2
17.Chao Dong, Yubin Deng, Chen Change Loy, and Xiaoou Tang. Compression artifacts reduction by a deep convolutional network. In IEEE International Conference on Computer Vision, pages 576–584, 2015. 2, 7
18.Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang. Learning a deep convolutional network for image super-resolution. In European Conference on Computer Vision, pages 184–199, 2014. 1, 2
19.Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2
20.Max Ehrlich, Larry Davis, Ser-Nam Lim, and Abhinav Shrivastava. Quantization guided jpeg artifact correction. In European Conference on Computer Vision, pages 293–309, 2020. 7
21.Gamaleldin Elsayed, Prajit Ramachandran, Jonathon Shlens, and Simon Kornblith. Revisiting spatial invariance with low-rank local connectivity. In International Conference on Machine Learning, pages 2868–2879, 2020. 2, 3
22.Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Pointwise shape-adaptive dct for high-quality denoising and deblocking of grayscale and color images. IEEE Transactions on Image Processing, 16(5):1395–1411, 2007. 7
23.Rich Franzen. Kodak lossless true color image suite. source: http://r0k. us/graphics/kodak, 4(2), 1999. 8
24.Manuel Fritsche, Shuhang Gu, and Radu Timofte. Frequency separation for real-world super-resolution. In IEEE Conference on International Conference on Computer Vision Workshops, pages 3599–3608, 2019. 5, 7
25.Xueyang Fu, Menglu Wang, Xiangyong Cao, Xinghao Ding, and Zheng-Jun Zha. A model-driven deep unfolding method for jpeg artifacts removal. IEEE Transactions on Neural Networks and Learning Systems, 2021. 2
26.Xueyang Fu, Zheng-Jun Zha, Feng Wu, Xinghao Ding, and John Paisley. Jpeg artifacts reduction via deep convolutional sparse coding. In IEEE International Conference on Computer Vision, pages 2501–2510, 2019. 2
27.Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems, pages 2672–2680, 2014. 3
28.Shuhang Gu, Nong Sang, and Fan Ma. Fast image super resolution via local regression. In IEEE Conference on International Conference on Pattern Recognition, pages 3128–3131, 2012. 1, 2
29.Shuhang Gu, Lei Zhang, Wangmeng Zuo, and Xiangchu Feng. Weighted nuclear norm minimization with application to image denoising. In IEEE conference on computer vision and pattern recognition, pages 2862–2869, 2014. 7, 8
30.Yong Guo, Jian Chen, Jingdong Wang, Qi Chen, Jiezhang Cao, Zeshuai Deng, Yanwu Xu, and Mingkui Tan. Closed-loop matters: Dual regression networks for single image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5407–5416, 2020. 2
31.Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Deep back-projection networks for super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1664–1673, 2018. 5, 6
32.Kaiming He, Jian Sun, and Xiaoou Tang. Single image haze removal using dark channel prior. IEEE transactions on Pattern Analysis and Machine Intelligence, 33(12):2341–2353, 2010. 2
33.Xiangyu He, Zitao Mo, Peisong Wang, Yang Liu, Mingyuan Yang, and Jian Cheng. Ode-inspired network design for single image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1732–1741, 2019. 6
34.Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5197–5206, 2015. 6, 8
35.Zheng Hui, Xinbo Gao, Yunchu Yang, and Xiumei Wang. Lightweight image super-resolution with information multi-distillation network. In ACM International Conference on Multimedia, pages 2024–2032, 2019. 5, 6
36.Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent structure-detail network. In European Conference on Computer Vision, pages 645–660. Springer, 2020. 2
37.Xiaozhong Ji, Yun Cao, Ying Tai, Chengjie Wang, Jilin Li, and Feiyue Huang. Real-world super-resolution via kernel estimation and noise injection. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 466–467, 2020. 5, 7
38.Xixi Jia, Sanyang Liu, Xiangchu Feng, and Lei Zhang. Focnet: A fractional optimal control network for image denoising. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6054–6063, 2019. 2, 7, 8
39.Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European Conference on Computer Vision, pages 694–711. Springer, 2016. 3
40.Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1646–1654, 2016. 1, 2, 6
41.Yoonsik Kim, Jae Woong Soh, Jaewoo Park, Byeongyong Ahn, Hyun-Seung Lee, Young-Su Moon, and Nam Ik Cho. A pseudo-blind convolutional neural network for the reduction of compression artifacts. IEEE Transactions on Circuits and Systems for Video Technology, 30(4):1121–1135, 2019. 2
42.Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 624–632, 2017. 2
43.Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photo-realistic single image super-resolution using a generative adversarial network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4681–4690, 2017. 1
44.Wenbo Li, Kun Zhou, Lu Qi, Nianjuan Jiang, Jiangbo Lu, and Jiaya Jia. Lapar: Linearly-assembled pixel-adaptive regression network for single image super-resolution and beyond. arXiv preprint arXiv:2105.10422, 2021. 5, 6
45.Yawei Li, Kai Zhang, Jiezhang Cao, Radu Timofte, and Luc Van Gool. Localvit: Bringing locality to vision transformers. arXiv preprint arXiv:2104.05707, 2021. 2
46.Zhen Li, Jinglei Yang, Zheng Liu, Xiaomin Yang, Gwanggil Jeon, and Wei Wu. Feedback network for image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3867–3876, 2019. 1
47.Dingkang Liang, Xiwu Chen, Wei Xu, Yu Zhou, and Xiang Bai. Transcrowd: Weakly-supervised crowd counting with transformer. arXiv preprint arXiv:2104.09116, 2021. 2
48.Jingyun Liang, Andreas Lugmayr, Kai Zhang, Martin Danelljan, Luc Van Gool, and Radu Timofte. Hierarchical conditional flow: A unified framework for image super-resolution and image rescaling. In IEEE Conference on International Conference on Computer Vision, 2021. 2
49.Jingyun Liang, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Mutual affine network for spatially variant kernel estimation in blind image super-resolution. In IEEE Conference on International Conference on Computer Vision, 2021. 2
50.Jingyun Liang, Kai Zhang, Shuhang Gu, Luc Van Gool, and Radu Timofte. Flow-based kernel prior with application to blind super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10601–10610, 2021. 2
51.Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 136–144, 2017. 1, 4, 6
52.Ding Liu, Bihan Wen, Yuchen Fan, Chen Change Loy, and Thomas S Huang. Non-local recurrent network for image restoration. arXiv preprint arXiv:1806.02919, 2018. 2, 7, 8
53.Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International Journal of Computer Vision, 128(2):261–318, 2020. 2
54.Pengju Liu, Hongzhi Zhang, Kai Zhang, Liang Lin, and Wangmeng Zuo. Multi-level wavelet-cnn for image restoration. In IEEE conference on computer vision and pattern recognition workshops, pages 773–782, 2018. 7, 8
55.Yun Liu, Guolei Sun, Yu Qiu, Le Zhang, Ajad Chhatkuli, and Luc Van Gool. Transformer in convolutional neural networks. arXiv preprint arXiv:2106.03180, 2021. 2
56.Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv preprint arXiv:2103.14030, 2021. 1, 2, 4
57.Xiaotong Luo, Yuan Xie, Yulun Zhang, Yanyun Qu, Cuihua Li, and Yun Fu. Latticenet: Towards lightweight image super-resolution with lattice block. In European Conference on Computer Vision, pages 272–289, 2020. 5, 6
58.David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In IEEE Conference on International Conference on Computer Vision, pages 416–423, 2001. 6
59.David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In IEEE International Conference on Computer Vision, pages 416–423, 2001. 8
60.Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia Tools and Applications, 76(20):21811–21838, 2017. 4, 5, 6
61.Yiqun Mei, Yuchen Fan, and Yuqian Zhou. Image super-resolution with non-local sparse attention. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3517–3526, 2021. 2, 5, 6
62.Tomer Michaeli and Michal Irani. Nonparametric blind super-resolution. In IEEE Conference on International Conference on Computer Vision, pages 945–952, 2013. 2
63.Ben Niu, Weilei Wen, Wenqi Ren, Xiangde Zhang, Lianping Yang, Shuzhen Wang, Kaihao Zhang, Xiaochun Cao, and Haifeng Shen. Single image super-resolution via a holistic attention network. In European Conference on Computer Vision, pages 191–207, 2020. 2, 4, 5, 6
64.Yali Peng, Lu Zhang, Shigang Liu, Xiaojun Wu, Yu Zhang, and Xili Wang. Dilated residual networks with symmetric skip connection for image denoising. Neurocomputing, 345:67–76, 2019. 2, 8
65.Tobias Plotz and Stefan Roth. Neural nearest neighbors networks. arXiv preprint arXiv:1810.12575, 2018. 7, 8
66.Prajit Ramachandran, Niki Parmar, Ashish Vaswani, Irwan Bello, Anselm Levskaya, and Jonathon Shlens. Stand-alone self-attention in vision models. arXiv preprint arXiv:1906.05909, 2019. 2
67.HR Sheikh. Live image quality assessment database release 2. http://live. ece. utexas. edu/research/quality, 2005. 7
68.Wenzhe Shi, Jose Caballero, Ferenc Huszar, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1874–1883, 2016. 3
69.Guolei Sun, Yun Liu, Thomas Probst, Danda Pani Paudel, Nikola Popovic, and Luc Van Gool. Boosting crowd counting with transformers. arXiv preprint arXiv:2105.10926, 2021. 2
70.Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In IEEE International Conference on Computer Vision, pages 4539–4547, 2017. 2
71.Chunwei Tian, Yong Xu, and Wangmeng Zuo. Image denoising using deep cnn with batch renormalization. Neural Networks, 121:461–473, 2020. 8
72.Radu Timofte, Vincent De Smet, and Luc Van Gool. Anchored neighborhood regression for fast example-based super-resolution. In IEEE Conference on International Conference on Computer Vision, pages 1920–1927, 2013. 2
73.Radu Timofte, Vincent De Smet, and Luc Van Gool. A+: Adjusted anchored neighborhood regression for fast super-resolution. In Asian Conference on Computer Vision, pages 111–126, 2014. 1, 2
74.Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve J egou. Training data-efficient image transformers & distillation through attention. arXiv preprint arXiv:2012.12877, 2020. 1, 2
75.Ashish Vaswani, Prajit Ramachandran, Aravind Srinivas, Niki Parmar, Blake Hechtman, and Jonathon Shlens. Scaling local self-attention for parameter efficient visual backbones. arXiv preprint arXiv:2103.12731, 2021. 2, 3
76.Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. arXiv preprint arXiv:1706.03762, 2017. 1, 2, 4
77.Longguang Wang, Yingqian Wang, Xiaoyu Dong, Qingyu Xu, Jungang Yang, Wei An, and Yulan Guo. Unsupervised degradation representation learning for blind super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10581–10590, 2021. 2
78.Longguang Wang, Yingqian Wang, Zhengfa Liang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning parallax attention for stereo image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12250–12259, 2019. 2
79.Longguang Wang, Yingqian Wang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning a single network for scale-arbitrary super-resolution. In IEEE Conference on International Conference on Computer Vision, pages 10581–10590, 2021. 2
80.Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. arXiv preprint arXiv:2107.10833, 2021. 3
81.Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In European Conference on Computer Vision Workshops, pages 701–710, 2018. 1, 2, 3, 5, 6, 7
82.Zhendong Wang, Xiaodong Cun, Jianmin Bao, and Jianzhuang Liu. Uformer: A general u-shaped transformer for image restoration. arXiv preprint arXiv:2106.03106, 2021. 2
83.Yunxuan Wei, Shuhang Gu, Yawei Li, Radu Timofte, Longcun Jin, and Hengjie Song. Unsupervised real-world image super resolution via domain-distance aware training. In IEEE Conference on Computer Vision and Pattern Recognition, pages 13385–13394, 2021. 2
84.Bichen Wu, Chenfeng Xu, Xiaoliang Dai, Alvin Wan, Peizhao Zhang, Zhicheng Yan, Masayoshi Tomizuka, Joseph Gonzalez, Kurt Keutzer, and Peter Vajda. Visual transformers: Token-based image representation and processing for computer vision. arXiv preprint arXiv:2006.03677, 2020. 2
85.Zhihao Xia and Ayan Chakrabarti. Identifying recurring patterns with deep neural networks for natural image denoising. In IEEE Winter Conference on Applications of Computer Vision, pages 2426–2434, 2020. 8
86.Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dollar, and Ross Girshick. Early convolutions help transformers see better. arXiv preprint arXiv:2106.14881, 2021. 2
87.Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In International Conference on Curves and Surfaces, pages 711–730, 2010. 6
88.Kai Zhang, Yawei Li, Wangmeng Zuo, Lei Zhang, Luc Van Gool, and Radu Timofte. Plug-and-play image restoration with deep denoiser prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 2, 7, 8
89.Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In IEEE Conference on International Conference on Computer Vision, 2021. 1, 3, 5, 7
90.Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142–3155, 2017. 1, 2, 7, 8
91.Kai Zhang, Wangmeng Zuo, Shuhang Gu, and Lei Zhang. Learning deep cnn denoiser prior for image restoration. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3929–3938, 2017. 1, 7, 8
92.Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn-based image denoising. IEEE Transactions on Image Processing, 27(9):4608–4622, 2018. 1, 2, 7, 8
93.Kai Zhang, Wangmeng Zuo, and Lei Zhang. Learning a single convolutional super-resolution network for multiple degradations. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3262–3271, 2018. 1, 2
94.Lei Zhang, Xiaolin Wu, Antoni Buades, and Xin Li. Color demosaicking by local directional interpolation and nonlocal adaptive thresholding. Journal of Electronic imaging, 20(2):023016, 2011. 8
95.Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In European Conference on Computer Vision, pages 286–301, 2018. 1, 2, 4, 5, 6
96.Yulun Zhang, Kunpeng Li, Kai Li, Bineng Zhong, and Yun Fu. Residual non-local attention networks for image restoration. arXiv preprint arXiv:1903.10082, 2019. 2, 6, 7, 8
97.Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2472–2481, 2018. 1, 2, 6
98.Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(7):2480–2495, 2020. 2, 7, 8
99.Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6881–6890, 2021. 2
100.Shangchen Zhou, Jiawei Zhang, Wangmeng Zuo, and Chen Change Loy. Cross-scale internal graph neural network for image super-resolution. arXiv preprint arXiv:2006.16673, 2020. 2, 5, 6

Citation

MLA

Liang, J., et al. “SwinIR: Image Restoration Using Swin Transformer”. 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), 2021, pp. 1833–44, https://doi.org/10.1109/ICCVW54120.2021.00210.

APA

Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., & Timofte, R. (2021). SwinIR: Image Restoration Using Swin Transformer. 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), 1833–1844. https://doi.org/10.1109/ICCVW54120.2021.00210

Chicago

Liang, J., J. Cao, G. Sun, K. Zhang, L. Van Gool, and R. Timofte. 2021. “SwinIR: Image Restoration Using Swin Transformer”. 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), 1833–44. https://doi.org/10.1109/ICCVW54120.2021.00210.

Harvard

Liang, J. et al. (2021) “SwinIR: Image Restoration Using Swin Transformer”, 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW). IEEE, pp. 1833–1844. Available at: https://doi.org/10.1109/ICCVW54120.2021.00210.

Vancouver

1. Liang J, Cao J, Sun G, Zhang K, Van Gool L, Timofte R (2021) SwinIR: Image Restoration Using Swin Transformer. In: 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW). IEEE, pp 1833–1844

BibTeX

@inproceedings{Liang_2021, title={SwinIR: Image Restoration Using Swin Transformer}, url={http://dx.doi.org/10.1109/ICCVW54120.2021.00210}, DOI={10.1109/iccvw54120.2021.00210}, booktitle={2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW)}, publisher={IEEE}, author={Liang, Jingyun and Cao, Jiezhang and Sun, Guolei and Zhang, Kai and Van Gool, Luc and Timofte, Radu}, year={2021}, month=Oct, pages={1833–1844} }

Metadata:Crossref

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: IEEE

SwinIR: Image Restoration Using Swin Transformer

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — SwinIR Overall Architecture for Image Restoration

Knowl 2 — Residual Swin Transformer Block (RSTB)

Knowl 3 — Swin Transformer Layer (STL) Formulation in SwinIR

Knowl 4 — Loss Functions for Restoration Objectives in SwinIR

Knowl 5 — Quantitative Evaluation on Classical Image Super-Resolution

Knowl 6 — Quantitative Evaluation on Lightweight Image Super-Resolution

Knowl 7 — Quantitative Evaluation on Grayscale and Color Image Denoising

Knowl 8 — Quantitative Evaluation on JPEG Compression Artifact Reduction

Knowl 9 — Ablation of RSTB Internal Design and Model Hyperparameters

Knowl 10 — Convergence Speed and Data Efficiency of SwinIR vs. CNNs

References

Citation

BibTeX

Source Code

Access the Paper