UCAN: Unified Convolutional Attention Network for Expansive Receptive Fields in Lightweight Super-Resolution

Cao Thien Tan; Phan Thi Thu Trang; Do Nghiem Duc; Ho Ngoc Anh; Hanyang Zhuang; Nguyen Duc Dung

cs.CVMar 12, 2026

UCAN: Unified Convolutional Attention Network for Expansive Receptive Fields in Lightweight Super-Resolution

Authors: Cao Thien Tan, Phan Thi Thu Trang, Do Nghiem Duc, Ho Ngoc Anh, Hanyang Zhuang, Nguyen Duc Dung

Organizations: Ho Chi Minh City Open University · AI Tech Lab, Ho Chi Minh City University of Technology · Code Mely AI Research Team · Ha Noi University of Science and Technology ·

Abstract

Hybrid CNN-Transformer architectures achieve strong results in image super-resolution, but scaling attention windows or convolution kernels significantly increases computational cost, limiting deployment on resource-constrained devices. We present UCAN, a lightweight network that unifies convolution and attention to expand the effective receptive field efficiently. UCAN combines window-based spatial attention with a Hedgehog Attention mechanism to model both local texture and long-range dependencies, and introduces a distillation-based large-kernel module to preserve high-frequency structure without heavy computation. In addition, we employ cross-layer parameter sharing to further reduce complexity. On Manga109 ( $4\times$ ), UCAN-L achieves 31.63 dB PSNR with only 48.4G MACs, surpassing recent lightweight models. On BSDS100, UCAN attains 27.79 dB, outperforming methods with significantly larger models. Extensive experiments show that UCAN achieves a superior trade-off between accuracy, efficiency, and scalability, making it well-suited for practical high-resolution image restoration.

Figures & tables

Figure 1: Performance comparison of PSNR versus model parameters on the Manga109 ( $\times$ 4) dataset. Our method is evaluated alongside state-of-the-art super-resolution approaches. Green triangles represent CNN-based methods, yellow squares denote Transformer-based methods, blue circles indicate SSM-based methods, and red diamonds show our proposed approach.

Figure 2: Comparison of feature maps output by Linear Attention (Using ReLU, ELU feature maps) and Hedgehog Attention. All experiments are conducted based on an image with $N$ = 256 and $d$ = 48. The full rank of matrices in the figure is 64. This figure shows that our model outperforms previous methods in diversity.

Figure 3: Detailed architecture of (a) Shared and Received Hybrid Attention (SHA and RHA) and (b) Large Kernel Distillation (LKD). LKD contains a Triple Feature Extraction block with three branches, which are the Large Kernel Spatial Branch, the Channel Branch, and the Small Kernel Spatial Branch. SHA and RHA employ Shared and Received Window Multi Head Self Attention (Shared WMHSA and Received WMHSA) to capture local information, and include a Shared Dual Fusion Layer (SDFL) and a Dual Fusion Receiver Layer (DFRL) to aggregate global context. The Dual Fusion Layer comprises two sub branches, which are Hedgehog Attention (HgA) and Channel Attention (CA).

				Set5		Set14		BSDS100		Urban100		Manga109
Method	scale	#param	MACs	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM
MambaIR-light [ 10 ]	$2\times$	905K	334.2G	38.13	0.9610	33.95	0.9208	32.31	0.9013	32.85	0.9349	39.20	0.9782
OmniSR [ 35 ]	$2\times$	772K	194.5G	38.22	0.9613	33.98	0.9210	32.36	0.9020	33.05	0.9363	39.28	0.9784
SRFormer-light [ 46 ]	$2\times$	853K	236.3G	38.23	0.9613	33.94	0.9209	32.36	0.9019	32.91	0.9353	39.28	0.9785
ATD-light [ 40 ]	$2\times$	753K	380.0G	38.29	0.9616	34.10	0.9217	32.39	0.9023	33.27	0.9375	39.52	0.9789
HiT-SRF [ 42 ]	$2\times$	847K	226.5G	38.26	0.9615	34.01	0.9214	32.37	0.9023	33.13	0.9372	39.47	0.9787

Table 1: Quantitative comparison on lightweight image super-resolution with state-of-the-art methods. The best and second-best results are shown in bold and underlined , respectively.

Panel of Figure 4 — Figure 4: Visual comparison between ground truth and different methods on Urban100.

Figure 5: Visualize in detail ERF of MambaIR [ 10 ] , MambaIRv2 [ 8 ] and UCAN

Panel of Figure 6 — Figure 6: Local attribution maps (LAM) comparison of different super-resolution methods at scale $\times$ 4.

Module	$\mathbb{R}^{64\times 64\times 64}$		$\mathbb{R}^{128\times 128\times 64}$
Module	Latency (ms)	#params	Latency (ms)	#params
Attention	$596.53\left(\text{\scriptsize 1.5×}\right)$	0.033M	$2576.75\left(\text{\scriptsize 1.0×}\right)$	0.082M
Flash Attention	$190.50\left(\text{\scriptsize 4.7×}\right)$		$191.80\left(\text{\scriptsize{\color[rgb]{1,0,0}13.4×}}\right)$
DFL	$903.47\left(\text{\scriptsize 1.0×}\right)$	0.014M	$1294.83\left(\text{\scriptsize 2.0×}\right)$	0.014M

Table 2: Comparison of Latency and Parameters between Naive Self-Attention, Flash Attention [ 3 ] , and the Dual Fusion Layer (DFL)

Block	Case	Set5		Urban100
Block	Case	PSNR	SSIM	PSNR	SSIM
High performance attention (HPA)	w/o HPA	38.27	0.9616	32.90	0.9346
High performance attention (HPA)	window $16\times 16$	38.32	0.9617	33.04	0.9364
WMHA	WS $8\times 8$	38.32	0.9617	33.02	0.9361
Dual Fusion Layer (DFL)	Using ReLU	38.33	0.9618	33.16	0.9374
Dual Fusion Layer (DFL)	Using ELU + 1	38.33	0.9618	33.16	0.9373

Table 3: Ablation study. We train all models on DIV2K for 400K iterations, and test on Set5 and Urban100 (

\times

2). The final result is shown in the last row.

$k_{\text{core}}$	Dilation $d$	$k_{\text{extra}}$	Final ERF
3	1	–	5
5	1	–	9
5	2	–	13
5	3	11	47
5	3	13	53
5	3	17	65

Table 4: LKSA module configurations and resulting effective kernel sizes (1D ERF). For rows with

k_{\text{extra}}=\text{--}

, we use the

\mathcal{F}_{LK\text{-S}}

formula ( 16 ); otherwise, we use the

\mathcal{F}_{LK\text{-L}}

formula ( 17 ).

Figure 7: Visualization of attention maps for Linear Attention using ReLU and $\text{ELU}+1$ (computed with sequence length $N=256$ ). The lack of normalization leads to high-magnitude artifacts.

Figure 8: Ranking consistency analysis. We compare the output ranking of Linear Attention using standard ReLU, Symmetric ReLU, and the Hedgehog Feature Map (sequence length $N=256$ ). While adding negative information (Sym-ReLU) improves consistency, Hedgehog achieves superior performance through learnable stability.

Case	Configuration	Set5		Urban100
Case	Configuration	PSNR	SSIM	PSNR	SSIM
Hybrid Block (HB)	depth=3	38.30	0.9617	33.05	0.9363
	depth=2	38.24	0.9615	32.95	0.9353
	depth=1	38.16	0.9612	32.71	0.9336
Large Kernel Distillation (LKD)	LKD depth=4	38.33	0.9618	33.12	0.9369
	LKD depth=3	38.33	0.9618	33.08	0.9366

Table 5: More ablation study. We train all models on DIV2K for 400K iterations, and test on Set5 and Urban100 (

\times 2

				Set5		Set14		BSDS100		Urban100		Manga109
Method	scale	#param	FLOPs	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM
SwinIR-light [ 19 ]	$2\times$	910K	244.2G	38.14	0.9611	33.86	0.9206	32.31	0.9012	32.76	0.9340	39.12	0.9783
ELAN [ 43 ]	$2\times$	621K	245.2G	38.27	0.9616	33.94	0.9207	32.30	0.9012	32.76	0.9340	39.11	0.9782
IPG-Tiny [ 34 ]	$2\times$	621K	203.1G	38.17	0.9611	34.24	0.9236	32.35	0.9018	33.04	0.9359	39.31	0.9786
PFT-light [ 23 ]	$2\times$	776K	278.3G	38.36	0.9620	34.19	0.9232	32.43	0.9030	33.67	0.9411	39.55	0.9792
UCAN (Our)	$2\times$	689K	146.3G	38.34	0.9618	34.27	0.9242	32.39	0.9025	33.22	0.9379	39.54	0.9790

Table 6: Quantitative comparison on lightweight image super-resolution with state-of-the-art methods. The best and second-best results are shown in bold and underlined , respectively.

Panel of Figure 9 — Figure 9: Visual comparison between the ground truth and different methods on Set5 - baby.

Panel of Figure 10 — Figure 10: Visual comparison between the ground truth and different methods on Set14 - man.

Panel of Figure 11 — Figure 11: Visual comparison between the ground truth and different methods on B100 - 300091.

Panel of Figure 12 — Figure 12: Visual comparison between the ground truth and different methods on Manga109 - Yumeko Cooking.

Panel of Figure 13 — Figure 13: Visual comparison between the ground truth and different methods on Manga109 - Gakuen Noise.

Panel of Figure 14 — Figure 14: Visual comparison between the ground truth and different methods on Manga109 - Yasakii Akuma.

Panel of Figure 15 — Figure 15: Visual comparison between the ground truth and different methods on Urban100 - 015.

Panel of Figure 16 — Figure 16: Visual comparison between the ground truth and different methods on Urban100 - img19.

Panel of Figure 17 — Figure 17: Visual comparison between the ground truth and different methods on Urban100 - img24.

Panel of Figure 18 — Figure 18: Visual comparison between the ground truth and different methods on Urban100 - img72.

Image Super-Resolution

Image super-resolution reconstructs high-resolution images from lower-resolution observations, balancing detail, fidelity, and computational cost.

See all 247 papers

Efficient Image Super-Resolution with Multi-Scale Spatial Adaptive Attention Networks
Feb 22, 2026Sushi Rao, Jingwei LiImage Super-Resolution Multi-Scale Attention
Zhejiang Gongshang University
CUST: Clustered Unit-level Similarity Transformer for Lightweight Image Super-Resolution
Jul 13, 2026Jeongsoo KimVision Transformer Efficient ViTs
Independent Researcher
SRGAN-CKAN: Expressive Super-Resolution with Nonlinear Functional Operators under Minimal Resources
May 2, 2026Roberto Isai Navaro-Aviña, Eduardo Said Merin-Martinez, Andres Mendez-Vazquez +1Kolmogorov-Arnold Networks Generative Adversarial Networks
Cinvestav, Unidad Guadalajara Av. del Bosque 1145, El Bajio Zapopan, Jalisco, 45017, Mexico · Cinvestav, Unidad Tamaulipas Km. 5.5 Carretera Victoria - Soto La Marina Victoria, 87139, Tamaulipas, Mexico

Feb 22, 2026J/K move · Enter open · S save

Efficient Image Super-Resolution with Multi-Scale Spatial Adaptive Attention Networks

Sushi Rao, Jingwei Li

Zhejiang Gongshang University

Image Super-Resolution Multi-Scale Attention Lightweight CNNs Multi-Scale Feature Fusion

Open paper AskPDF arXiv

This paper introduces a lightweight image super-resolution (SR) network, termed the Multi-scale Spatial Adaptive Attention Network (MSAAN), to address the common dilemma between high reconstruction fidelity and low model complexity in existing SR methods. The core of our approach is a novel Multi-scale Spatial Adaptive Attention Module (MSAA), designed to jointly model fine-grained local details and long-range contextual dependencies. The MSAA comprises two synergistic components: a Global Feature Modulation Module (GFM) that learns coherent texture structures through differential feature extraction, and a Multi-scale Feature Aggregation Module (MFA) that adaptively fuses features from local to global scales using pyramidal processing. To further enhance the network's capability, we propose a Local Enhancement Block (LEB) to strengthen local geometric perception and a Feature Interactive Gated Feed-Forward Module (FIGFF) to improve nonlinear representation while reducing channel redundancy. Extensive experiments on standard benchmarks (Set5, Set14, B100, Urban100, Manga109) across

\times2

\times3

, and

\times4

scaling factors demonstrate that both our lightweight (MSAAN-light) and standard (MSAAN) versions achieve superior or competitive performance in terms of PSNR and SSIM, while maintaining significantly lower parameters and computational costs than state-of-the-art methods. Ablation studies validate the contribution of each component, and visual results show that MSAAN reconstructs sharper edges and more realistic textures.

Abstract

Image Super-Resolution

Explore similar work

Efficient Image Super-Resolution with Multi-Scale Spatial Adaptive Attention Networks

CUST: Clustered Unit-level Similarity Transformer for Lightweight Image Super-Resolution

SRGAN-CKAN: Expressive Super-Resolution with Nonlinear Functional Operators under Minimal Resources