Perceptual Quality Optimization of Image Super-Resolution

Wei Zhou; Yixiao Li; Hadi Amirpour; Xiaoshuai Hao; Jiang Liu; Peng Wang; Hantao Liu

eess.IVFeb 25, 2026

Perceptual Quality Optimization of Image Super-Resolution

Authors: Wei Zhou, Yixiao Li, Hadi Amirpour, Xiaoshuai Hao, Jiang Liu, Peng Wang, Hantao Liu

Organizations: School of Computer Science and Informatics, Cardiff University, UK · Christian Doppler Laboratory ATHENA, Alpen-Adria-Universität, Klagenfurt, Austria · XiaomiEV, China · Centre for Vision, Speech and Signal Processing, University of Surrey, UK

Abstract

Single-image super-resolution (SR) has achieved remarkable progress with deep learning, yet most approaches rely on distortion-oriented losses or heuristic perceptual priors, which often lead to a trade-off between fidelity and visual quality. To address this issue, we propose an Efficient Perceptual Bi-directional Attention Network (Efficient-PBAN) that explicitly optimizes SR towards human-preferred quality. Unlike patch-based quality models, Efficient-PBAN avoids extensive patch sampling and enables efficient image-level perception. The proposed framework is trained on our self-constructed SR quality dataset that covers a wide range of state-of-the-art SR methods with corresponding human opinion scores. Using this dataset, Efficient-PBAN learns to predict perceptual quality in a way that correlates strongly with subjective judgments. The learned metric is further integrated into SR training as a differentiable perceptual loss, enabling closed-loop alignment between reconstruction and perceptual assessment. Extensive experiments demonstrate that our approach delivers superior perceptual quality. Code is publicly available at https://github.com/Lighting-YXLI/Efficient-PBAN.

Figures & tables

Figure 1: Overall framework of the proposed method. Left: SR optimization pipeline driven by Efficient-PBAN. Right: Efficient-PBAN structure with two-stage training, where the “ice” and “fire” icons denote frozen and trainable parameters, respectively. And the details of PBA + block are described in Section 2.1 . $\mathcal{L}_{D}$ and $\mathcal{L}_{P}$ refer to the distortion-oriented and perceptual loss, respectively.

Test	SR	Model	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PFIQA $\uparrow$	EPBAN $\uparrow$	SF $\uparrow$	SN $\uparrow$	SFSN $\uparrow$
[ 12 ]	[ 16 ]	Original	26.3513	0.7188	0.3382	3.8528	2.8682	0.9857	2.1952	1.1067
		+SSIM	25.6755	0.7304	0.3358	3.8702	2.9415	0.9764	2.3886	1.1177
		+EPBAN	25.6203	0.6963	0.3567	3.8911	3.3165	0.9836	2.4982	1.1351
		+SSIM $\&$ EPBAN	25.6741	0.7110	0.3560	3.9348	3.4025	0.9843	2.4084	1.1268
[ 1 ]		Original	29.2358	0.8251	0.2928	3.6466	2.9568	0.9907	1.9200	1.0836
		+SSIM	28.4237	0.8309	0.2931	3.6586	2.9998	0.9861	2.0532	1.0928

Table 1: Quantitative comparison based on CAMixerSR [ 16 ] and LINF [ 18 ] across B100 [ 12 ] and DIV2K [ 1 ] databases with different optimization counterparts. Results are reported as mean score on each testing database. Best and second-best results for each database and baseline are highlighted in red and blue , respectively. Note that EPBAN refers to the Efficient-PBAN.

Figure 2: Visualizations of the original SR images and the corresponding optimization results, where the regions highlighted by red boxes are enlarged for clearer illustration.

$\beta$ / $\alpha$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PFIQA $\uparrow$	Efficient-PBAN $\uparrow$	SF $\uparrow$	SN $\uparrow$	SFSN $\uparrow$
1/9	26.1260	0.7195	0.3315	3.8717	2.6703	0.9848	2.1308	1.0994
2/8	26.1273	0.7195	0.3314	3.8716	2.6707	0.9848	2.1302	1.0993
3/7	26.1292	0.7194	0.3314	3.8710	2.6712	0.9848	2.1292	1.0992
4/6	26.1334	0.7192	0.3313	3.8692	2.6725	0.9849	2.1269	1.0991
5/5	25.6741	0.7110	0.3560	3.9348	3.4025	0.9843	2.4084	1.1268
6/4	20.2765	0.3358	0.5914	4.9893	4.3056	0.9699	3.4414	1.2171

Table 2: Ablation study on the weighting ratio between the distortion-oriented loss

\mathcal{L}_{D}

(denoted as

\alpha

) and the perceptual loss

\mathcal{L}_{P}

(denoted as

\beta

) for CAMixerSR on the B100 dataset. The best performances are highlighted in bold.

Explore similar work

FreqOrtho-SR: Frequency-Guided Orthogonal Expert Learning for Real-World Image Super-Resolution
Jun 27, 2026Minh Son Hoang, Dinh Phu Tran, Quyen Nguyen Duc +2Diffusion Models for Image Restoration Mixture of Experts
School of Computing, KAIST, Republic of Korea
Interest Entanglement: The Hidden Barrier to Blind Super-Resolution Optimization
Jun 21, 2026Junxiong Lin, Xinji Mai, Qianyu Guo +5Multi-Objective Optimization Image Super-Resolution
College of Intelligent Robotics and Advanced Manufacturing, Fudan University · Shanghai Institute of Virology, Shanghai Jiao Tong University School of Medicine · The University of Hong Kong +1
Single Image Super-Resolution via Bivariate `A Trous Wavelet Diffusion
Mar 7, 2026Maryam Heidari, Nantheera Anantrasirichai, Alin AchimImage Super-Resolution Diffusion Models for Image Restoration
Visual Information Laboratory, University of Bristol, Bristol, UK

Jun 27, 2026J/K move · Enter open · S save

FreqOrtho-SR: Frequency-Guided Orthogonal Expert Learning for Real-World Image Super-Resolution

Minh Son Hoang, Dinh Phu Tran, Quyen Nguyen Duc, Dam Hoang Phuong+1

School of Computing, KAIST, Republic of Korea

Diffusion Models for Image Restoration Mixture of Experts Image Super-Resolution

Open paper AskPDF arXiv

Diffusion prior-based methods have shown impressive results in real-world image super-resolution (ISR), yet two key challenges persist: balancing pixel-level fidelity with semantic quality, and adapting to diverse degradations. Existing dual-branch approaches freeze the pixel module during semantic training, but the semantic branch can still expand capacity within the pixel subspace, precluding genuine perceptual improvement. Moreover, using a single static adapter cannot generalize across heterogeneous real-world corruptions. To address both issues, we propose FreqOrtho-SR, which comprises:

\textbf{Freq}

uency-guided Mixture of LoRA Experts (FreqMoE), it routes inputs to specialized experts via a non-parametric FFT-based degradation-feature extractor that encodes frequency-domain signatures, enabling stable and interpretable specialization across corruption types; and

\textbf{Ortho}

gonal Gradient Projection (OGP), which reframes the dual-objective optimization as a subspace-constrained problem: by extracting the pixel-fidelity subspace via SVD on combined expert weight deltas and projecting semantic gradients onto its null space, OGP guarantees orthogonality between the two objectives, enabling genuinely complementary learning without mutual interference. Experiments show that FreqOrtho-SR achieves competitive overall performance and a strong fidelity-perception trade-off across multiple benchmarks with efficient single-step inference. The source code of our method can be found at sonhm3029/FreqOrtho-SR.

Abstract

Explore similar work

FreqOrtho-SR: Frequency-Guided Orthogonal Expert Learning for Real-World Image Super-Resolution

Interest Entanglement: The Hidden Barrier to Blind Super-Resolution Optimization

Single Image Super-Resolution via Bivariate `A Trous Wavelet Diffusion