Learning Polarization Image Restoration with General Restoration Priors

Chenggong Li; Jinhao Liu; Caiyun Wu; Yidong Luo; Junchao Zhang; Degui Yang

cs.CVSep 25, 2026

Learning Polarization Image Restoration with General Restoration Priors

Authors: Chenggong Li, Jinhao Liu, Caiyun Wu, Yidong Luo, Junchao Zhang, Degui Yang

Organizations: School of Automation Central South University · School of Engineering Westlake University

Abstract

Polarization imaging captures distinctive surface and geometric cues that benefit a wide range of vision tasks. However, real-world polarization acquisition is often affected by multiple coupled degradations, making image restoration essential for practical polarization vision. Existing methods are largely tailored to specific degradations and remain constrained by the limited scale and quality of polarization data. To address these limitations, we develop an all-in-one polarization restoration framework for diverse and composite degradations. We first study the impact of different polarization representations on restoration performance and identify the normalized Stokes representation as an effective choice for separating intensity and polarization information. Accordingly, we devise a dual-branch architecture that separates intensity and polarization modeling. To overcome the limitations of polarization-specific training, the intensity branch leverages pretrained general restoration priors and a mixture-of-experts extension for composite degradations, while its restoration knowledge is adaptively distilled into the symmetric polarization branch via a cross-domain feature transform. In addition, we establish a composite-degradation polarization benchmark to support all-in-one restoration research. Extensive experiments on public datasets and our proposed benchmark demonstrate the effectiveness of the proposed method.

Figures & tables

Figure 1: Polarization restoration approach. (a) Angle-wise general restoration overlooks polarization correlations. (b) Degradation-specific experts (“Polar Model”) handle only individual degradations and remain constrained by polarization data distribution. (c) Our framework combines general restoration priors with polarization-specific modeling for restoration across diverse degradations. (d) Visual comparisons on blur (B), low light (L), mosaic (M), and their combinations: low quality image (LQ), MoCE-IR ( Zamfir et al. (2025) ), SSMD ( Zhou et al. (2026) ), Ours, and ground truth (GT) from left to right.

Figure 2: Representation study. (a) Representation visualization. (b) DoLP metric across intensity levels. Higher is better. (c) AoP metric across DoLP levels. Lower is better.

Type	$PSNR_{I}$ $\uparrow$	$PSNR_{p}$ $\uparrow$	$SSIM_{I}$ $\uparrow$	$SSIM_{p}$ $\uparrow$	MAE $\downarrow$	$MAE_{r}$ $\downarrow$
(a)	30.489	29.390	0.864	0.713	16.541	12.973
(b)	30.603	29.642	0.864	0.723	15.740	12.618
(c)	30.515	30.110	0.862	0.727	15.603	12.101
(d)	30.118	29.424	0.857	0.712	16.982	13.852
(e)	30.268	29.790	0.858	0.722	14.987	12.503

Table 1: Average performance of different representations on three restoration tasks. From (a) to (e):

[x_{0^{\circ}},x_{45^{\circ}},x_{90^{\circ}},x_{135^{\circ}}]

[S_{0},S_{1},S_{2}]

[S_{0},\tilde{S}_{1},\tilde{S}_{2}]

[S_{0},p,\phi]

and

[S_{0},p,\cos(2\phi),\sin(2\phi)]

. The subscripts

I

and

p

denote intensity and DoLP, respectively.

MAE_{r}

is the AOP error in reliable regions with GT DoLP

>0.1

Figure 3: Overview of the proposed method.(a) The normalized Stokes representation is processed by a U-Net composed of intensity-polarization restoration blocks (IPRBs), with restoration behavior conditioned by textual guidance through instruction condition block (ICB ( Conde et al. (2024) )) . (b) Each IPRB contains intensity and polarization branches built upon NAFNet-style attention and feed-forward blocks ( Chen et al. (2022) ). Residual connections are employed but omitted in the figure for clarity. The intensity branch incorporates pretrained restoration priors ( Conde et al. (2024) ) and an MoE extension, while the polarization branch separately models two polarization components and aggregates them during attention. Finally, the intensity-polarization adaptive transfer module (IP-ATM) enables domain interaction and adapts restoration priors to the polarization domain.

Figure 4: IP-ATM mechanism. Polarization representations selectively incorporate intensity restoration priors through adaptive scaling and shifting, followed by zero-convolution modulation that feeds polarization cues back into the intensity features.

Method	$PSNR_{I}$ $\uparrow$	$PSNR_{p}$ $\uparrow$	$SSIM_{I}$ $\uparrow$	$SSIM_{p}$ $\uparrow$	MAE $\downarrow$
PolDeblur ∗	29.193	27.232	0.864	0.694	15.803
SSMD ∗	32.601	29.550	0.909	0.744	14.416
PromptIR	25.098	27.554	0.802	0.690	15.858
InstructIR	29.128	28.388	0.873	0.709	15.596
AdaIR	25.106	27.516	0.801	0.690	15.874
MoCE-IR	31.451	29.012	0.902	0.729	15.184

Table 2: Quantitative comparison under the AIO3 setting. The superscript

*

indicates task-specific weights. Bold indicates the best performance.

Method	$PSNR_{I}$ $\uparrow$	$PSNR_{p}$ $\uparrow$	$SSIM_{I}$ $\uparrow$	$SSIM_{p}$ $\uparrow$	MAE $\downarrow$
PromptIR	34.248	30.627	0.919	0.757	16.914
InstructIR	33.237	31.309	0.923	0.779	17.146
AdaIR	34.729	31.185	0.925	0.774	15.687
MoCE-IR	34.542	32.250	0.938	0.800	16.615
CDIR	34.112	30.945	0.924	0.771	16.173
SSMD	34.455	32.031	0.930	0.804	15.525

Table 3: Quantitative comparison on the proposed PolarCDD and average performance. Bold indicates the best performance.

Figure 5: Visual comparisons of different methods on PolarCDD, including PromptIR ( Potlapalli et al. (2023) ), InstructIR ( Conde et al. (2024) ), AdaIR ( Cui et al. (2025) ), CDIR ( Cui et al. (2026) ), MoCE-IR ( Zamfir et al. (2025) ), SSMD ( Zhou et al. (2026) ) and Ours. N is noise, B is blur, L is low light, and M is mosaic.

Configuration	$PSNR_{I}$ $\uparrow$	$PSNR_{p}$ $\uparrow$	$SSIM_{I}$ $\uparrow$	$SSIM_{p}$ $\uparrow$	MAE $\downarrow$
No promotion	36.649	31.125	0.945	0.757	14.316
$+$ Polarization-to-intensity	36.942	31.041	0.945	0.755	14.386
$+$ Intensity-to-polarization (ours)	37.518	32.820	0.949	0.794	13.009
IP-ATM $\rightarrow$ Zero-Conv	36.365	31.852	0.940	0.772	13.620
W/o gate	37.449	32.621	0.947	0.791	13.100
W/o MoE	35.610	32.319	0.938	0.786	13.362

Table 4: Ablation study of the proposed framework under the AIO3 setting.

Figure 6: Feature visualization of IP-ATM at the 8th layer under different degradations.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Component	Layer	Configuration	Output size
Text encoder	Language model	BGE-Micro-v2; mean pooling; frozen	$384$
	Projection head	Linear $384\!\rightarrow\!256$ ; $\ell_{2}$ normalization	$256$
Input	Representation	$[S_{0},\tilde{S}_{1},\tilde{S}_{2}]$ ; $3+3+3$ channels	$h\times w\times 9$
	Branch split	$S_{0}$ : 3 channels; $[\tilde{S}_{1},\tilde{S}_{2}]$ : 6 channels	$h\times w\times(3+6)$
	Intro	$3\times 3$ Conv for $S_{0}$ ; grouped $3\times 3$ Conv ( $g=2$ ) for $\tilde{S}_{1}/\tilde{S}_{2}$	$h\times w\times(32+32)$
Encoder 1	IPRB	$(B_{0},B_{12})=(2,1)$ ; $(c_{0},c_{12})=(32,32)$ ; ICB; IP-ATM	$h\times w\times(32+32)$

Appendix

Table 5: Configurations of the proposed network architecture.

B_{0}

and

B_{12}

denote the numbers of NAFBlocks in the intensity (

S_{0}

) and polarization (

\tilde{S}_{1}/\tilde{S}_{2}

) branches, respectively.

c_{0}

and

c_{12}

denote their feature-channel dimensions. ICB is the instruction condition block, IP-ATM is the intensity–polarization adaptive transfer module, and MoE uses

E

experts with top-

k

routing.

Figure 7: Degradation synthesis pipeline. low light: (a), blur: (b), noise: (c) or (d), mosaic: (e), low light+blur: (a)+(b), low light+mosaic: (a)+(e), blur+noise: (b)+(c), blur+mosaic: (b)+(e), noise+mosaic: (c)+(e) or (d)+(e), low light+blur+mosaic: (a)+(b)+(e), blur+noise+mosaic: (b)+(c)+(e).

Method	Params (M)	GPU memory (GB)	FLOPs (T)	Latency (ms)
NAFNet	17.112	1.068	0.294	82.91
NAFNet-9ch	17.115	0.321	0.074	29.14
Restormer	11.740	5.998	1.029	684.08
Restormer-9ch	11.745	1.533	0.259	175.47
PromptIR	35.377	9.179	2.530	915.41
PromptIR-9ch	35.385	2.474	0.635	228.93

Appendix

Table 6: Complexity comparison. “9ch” denotes evaluation with a nine-channel normalized Stokes input adaptation. GPU memory, FLOPs, and latency are measured on

512\times 512

images using an NVIDIA RTX 4090 GPU.

Figure 8: Visual comparison on real-world motion blur. The superscript $\#$ denotes the pretrained model.

Configuration	$PSNR_{I}$ $\uparrow$	$PSNR_{p}$ $\uparrow$	$SSIM_{I}$ $\uparrow$	$SSIM_{p}$ $\uparrow$	MAE $\downarrow$
NAFNet-9ch	25.788	27.869	0.818	0.696	15.948
NAFNet	33.180	28.554	0.914	0.728	15.566
NAFNet-Ours	35.468	30.626	0.944	0.771	13.577
Restormer-9ch	33.307	29.536	0.923	0.743	14.281
Restormer	32.299	29.190	0.903	0.736	15.069
Restormer-Ours	36.574	31.014	0.951	0.780	13.246

Appendix

Table 7: Comparison of different input and branch designs with general restoration priors. “9ch” directly processes the nine-channel input, “3ch” uses the three-channel intensity representation, and “Ours” uses our intensity–polarization dual-branch design. We report the performance of all experiments at 100k training iterations. Bold indicates the best performance in each subtable.

Configuration	$PSNR_{I}$ $\uparrow$	$PSNR_{p}$ $\uparrow$	$SSIM_{I}$ $\uparrow$	$SSIM_{p}$ $\uparrow$	MAE $\downarrow$
Full polarization branch	36.034	32.342	0.934	0.784	13.325
Full branch w/o IP-ATM	35.141	31.119	0.929	0.756	14.323
Tiny polarization branch	35.744	32.017	0.934	0.777	13.573
Half polarization branch (Ours)	35.893	32.261	0.934	0.782	13.358

Appendix

Table 8: Ablation study on the polarization branch and feature promotion under the AIO3 setting. We report the performance of all experiments at 100k training iterations. Bold indicates the best performance.

Configuration	$PSNR_{I}$ $\uparrow$	$PSNR_{p}$ $\uparrow$	$SSIM_{I}$ $\uparrow$	$SSIM_{p}$ $\uparrow$	MAE $\downarrow$
$\mathcal{L}_{ang}$	36.259	31.224	0.937	0.757	14.986
$+$ $\mathcal{L}_{stokes}$	36.129	31.543	0.936	0.766	14.518
$+$ $\mathcal{L}_{p,\phi}$ (Ours)	35.893	32.261	0.934	0.782	13.358
w/o $\mathcal{L}_{ang}$	35.120	32.211	0.929	0.782	13.357

Appendix

Table 9: Ablation study on different loss-function configurations under the AIO3 settings. We report the performance of all experiments at 100k training iterations. Bold indicates the best performance.

Method	$PSNR_{I}$ $\uparrow$	$PSNR_{p}$ $\uparrow$	$SSIM_{I}$ $\uparrow$	$SSIM_{p}$ $\uparrow$	MAE $\downarrow$
Ours-noMoE	32.032	29.470	0.919	0.739	14.241
Ours	34.029	29.939	0.933	0.749	13.991

Appendix

Table 10: Comparison of Ours and Ours w/o MoE under the AIO3 setting. Bold indicates the best performance.

Figure 9: Activation patterns of each expert on PolarCDD. The probability indicates the proportion of images in a given task that are routed to the corresponding expert.

Figure 10: Visual comparisons of different methods on the dataset LLCP ( Xu et al. [2022] ) and PolarCDD. L is low-light. The superscript * indicates task-specific weights.

Figure 11: Visual comparisons of different methods on the dataset PolDeblur ( Zhou et al. [2025a] ) and real-world dataset PV ( Li et al. [2026b] ). B is blur. The superscript * indicates task-specific weights.

Figure 12: Visual comparisons of different methods on the dataset PolarCDD and real-world dataset ( Luo et al. [2025] ). N is noise and M is mosaic. The superscript * indicates task-specific weights.

Figure 13: Visual comparisons of different methods on the dataset PolarCDD. B is blur, L is low light, N is noise and M is mosaic.

Figure 14: Text-conditioned restoration behavior.

Explore similar work

PolUR: Polarization Aware Multimodal Degradation Perception for Underwater Image Restoration
Oct 13, 2026Wenshuo Zhang, Xiangyue Zhang, Junlin Li +5Image Restoration Underwater Image Enhancement
Faculty of Robot Science and Engineering, Northeastern University, Shenyang 110819, China
Architectural Unification for Polarimetric Imaging Across Multiple Degradations
Mar 6, 2026Chu Zhou, Yufei Han, Junda Liao +6Polarimetric Imaging Image Restoration
Digital Content and Media Sciences Research Division, National Institute of Informatics, Tokyo 101-8430, Japan. · Pattern Recognition and Intelligent System Laboratory, School of Artificial Intelligence, Beijing University of Posts and Telecommunications, Beijing 100876, China · State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University, Beijing 100080, China, and also with the National Engineering Research Center of Visual Technology, School of Computer Science, Peking University, Beijing 100080, China.
PolarAPP: Beyond Polarization Demosaicking for Polarimetric Applications
Mar 24, 2026Yidong Luo, Chenggong Li, Yunfeng Song +4Multi-Task Learning Polarimetric Imaging
Zhejiang University, Hangzhou, China · School of Automation, Central South University, Changsha, China · School of Engineering, Westlake University, Hangzhou, China +1

Oct 13, 2026J/K move · Enter open · S save

PolUR: Polarization Aware Multimodal Degradation Perception for Underwater Image Restoration

Wenshuo Zhang, Xiangyue Zhang, Junlin Li, Wenjie Tian+4

Faculty of Robot Science and Engineering, Northeastern University, Shenyang 110819, China

Image Restoration Underwater Image Enhancement All-in-One Image Restoration

Open paper AskPDF arXiv

Recovering high-quality images from degraded underwater observations is crucial for underwater exploration. Polarization observations show promise as an effective source of complementary physical cues for recovering target radiance under clear-water conditions. However, existing methods have yet to jointly model raw intensity information and polarization cues to enable image restoration driven by an understanding of mixed degradations (color casts, haze, and low light). To overcome these limitations, we propose PolUR, a unified multimodal language-driven framework that jointly uses polarization and RGB information to identify degradation types and severity levels through structured semantic instructions and enhances joint restoration representations under the guidance of degradation-aware semantic features to improve restoration robustness under mixed degradations. We further construct a new benchmark named MixPolar, which contains underwater images exhibiting real-world degradations and their corresponding degradation descriptions. MixPolar supports systematic evaluation of degradation perception, image restoration, and performance on downstream tasks. Extensive experiments demonstrate that PolUR effectively integrates polarization and RGB information, achieving competitive performance in degradation perception and image restoration, while also demonstrating generalization capability and utility for downstream tasks.

Abstract

Explore similar work

PolUR: Polarization Aware Multimodal Degradation Perception for Underwater Image Restoration

Architectural Unification for Polarimetric Imaging Across Multiple Degradations

PolarAPP: Beyond Polarization Demosaicking for Polarimetric Applications