UniPolar: Joint Image and Polarimetric Restoration under Compound Degradations via Physical Consistency Learning

Yang Jin; Guangyu Guo; Linghao Shen; Binglu Wang

cs.CVOct 16, 2026

UniPolar: Joint Image and Polarimetric Restoration under Compound Degradations via Physical Consistency Learning

Authors: Yang Jin, Guangyu Guo, Linghao Shen, Binglu Wang

Organizations: School of Astronautics, Northwestern Polytechnical University, Xi’an 710072, China · School of Computer Science and Technology, Zhejiang University, Hangzhou 310058, China · School of Marine Science and Technology, Tianjin University, Tianjin, 300072, China

Abstract

Polarimetric image restoration aims to recover polarization angle images and reliable polarimetric quantities from degraded observations. Beyond visual fidelity, it also requires preserving the physical relationships underlying reliable polarimetric estimation. However, existing methods typically address either device- or environment-induced degradations in isolation, limiting their applicability to color polarization filter array (CPFA) imaging, where both often coexist. Such compound degradations jointly perturb image intensities and cross angle polarization differences. Consequently, improved image fidelity alone does not necessarily ensure accurate Stokes, DoLP, and AoLP recovery. To address this challenge, we propose UniPolar, a coupled Image Stokes framework for joint image polarimetric restoration from compound-degraded CPFA observations. Specifically, UniPolar employs two complementary streams to reconstruct four polarization-angle images and directly predict Stokes difference components. To coordinate these streams, Degradation Conditioned Cross Domain Interaction combines Dual Domain Frequency Interaction for frequency selective evidence exchange with Degradation Conditioned Modulation for condition adaptive feature correction. Furthermore, at the output level, Image Stokes Physical Consistency Learning aligns image derived and directly predicted Stokes components while regularizing angular fidelity and the physical feasibility of DoLP. On the compound-degradation benchmark, UniPolar achieves average image and DoLP PSNR of 31.522 dB and 24.097 dB, respectively, with an average AoLP error of 16.833 degree. These results underscore the importance of jointly considering visual fidelity and polarimetric reliability, offering a physically grounded perspective on learning based restoration under compound degradations.

Figures & tables

Fig. 1: (a) Coupled device- and environment-induced degradations in CPFA observations impair both image and polarimetric restoration. (b) UniPolar jointly recovers polarization-angle images and polarimetric quantities, promoting Image–Stokes consistency from cross-domain representations to physically constrained outputs.

Fig. 2: Overview of UniPolar. Given compound-degraded CPFA samples and their masks, the image stream reconstructs four polarization-angle images, while the Stokes stream estimates the difference components $s_{1}$ and $s_{2}$ from degraded polarimetric cues. DDFI exchanges complementary low- and high-frequency information between the streams, and DCM adapts the coupled features to the degradation condition. Image–Stokes PCL finally reconciles image-derived and directly predicted Stokes quantities through component, angular, and DoLP-feasibility constraints, yielding visually faithful images and physically reliable polarimetric outputs.

Task-Specific Polarimetric Methods
Method	Normal			Low-light			Haze			Average
	Mean $\uparrow$	$p$ $\uparrow$	$\theta$ $\downarrow$	Mean $\uparrow$	$p$ $\uparrow$	$\theta$ $\downarrow$	Mean $\uparrow$	$p$ $\uparrow$	$\theta$ $\downarrow$	Mean $\uparrow$	$p$ $\uparrow$	$\theta$ $\downarrow$
TCPDNet [ 9 ]	36.790/0.930	27.721/0.587	19.152	30.836/0.828	18.927/0.437	32.659	14.190/0.499	16.534/0.291	29.391	27.272/0.752	21.061/0.438	27.067
PIDSR [ 16 ]	37.924 /0.939	28.194 /0.602	19.969	32.308/0.881	20.431/0.499	27.696	15.272/0.461	13.691/0.222	30.039	28.501/0.760	20.772/0.441	25.901
CPDDNet [ 14 ]	36.085/0.921	26.930/0.554	22.087	29.480/0.795	20.170/0.395	42.452	14.324/0.507	16.614/0.299	28.080	26.630/0.741	21.238/0.416	30.873
General Multi-Degradation Image Restoration Methods

TABLE I: Quantitative comparison of task-specific polarimetric restoration, general multi-degradation restoration, and joint image–polarimetric restoration methods for compound-degraded CPFA observations.

p

and

\theta

denote DoLP and AoLP, respectively. Top-three results are highlighted in red (1st), blue (2nd), and green (3rd). For input-matched comparison,

\ast

denotes the variants that use the same observation-derived degraded Stokes cues as UniPolar.

\dagger

indicates initialization with weights from denoising-and-demosaicking training on synthetic data [ 17 ] .

Fig. 3: t-SNE visualization of degradation representations. Blue , green and pink denote normal, low-light, and haze conditions, respectively.

Low-Light Enhancement
Method		Mean $\uparrow$	$s_{0}$ $\uparrow$	$p$ $\uparrow$	$\theta$ $\downarrow$
PLIE	PLIE [ 11 ]	-	39.1900 / 0.9770	27.1500 / 0.7650	15.9859
PLIE	Ours	36.0689 / 0.9401	37.0561 / 0.9625	22.4486 / 0.6765	14.0609
Denoising and Demosaicking
PIDSR	Polanalyser [ 59 ]	32.1525/0.8971	33.2800/0.9138	26.6800/0.7164	17.8666
	IGRI2 [ 35 ]	34.4200/0.9361	35.5000/0.9450	27.7800/0.7486	16.5830

TABLE II: Comparison with task-specific polarimetric restoration methods under individual device-based and environment-based degradation settings.

p

and

\theta

denote DoLP and AoLP, respectively. Top-three results are highlighted in red (1st), blue (2nd), and green (3rd).

Method		Mean $\uparrow$	$s_{0}$ $\uparrow$	$p$ $\uparrow$	$\theta$ $\downarrow$
Synthetic MQ	Polanalyser [ 59 ]	38.2149/0.9653	35.2560/0.9541	32.5532/0.8387	14.8220
	NLCSR [ 37 ]	36.8877/0.9523	33.5610/0.9310	34.0964/0.8557	11.1297
	CPDNet [ 7 ]	39.4374/0.9700	36.6062/0.9631	32.6586/0.8465	12.5138
	TCPDNet [ 9 ]	42.5602 /0.9815	39.7694/0.9765	36.0613/0.8921	10.1765
	DCPM [ 17 ]	42.0221/0.9807	40.0263 /0.9759	36.1496/ 0.8958	11.7376
	PIDSR [ 16 ]	42.4990/ 0.9822	40.6269 / 0.9794	36.5413 /0.8950	9.6463

TABLE III: Comparison with task-specific methods for joint denoising and demosaicking on Synthetic MQ and Synthetic PIDSR [ 17 ] .

p

and

\theta

denote DoLP and AoLP, respectively. Top-three results are highlighted in red (1st), blue (2nd), and green (3rd).

DD	DDFI	DCM	PCL	Mean $\uparrow$	$p$ $\uparrow$	$\theta$ $\downarrow$
				27.7079/0.7987	22.0222/0.4899	24.6531
$\checkmark$				28.9621/0.8338	23.1951/0.5281	20.9951
$\checkmark$	$\checkmark$			29.4692/0.8438	24.0186/0.5588	19.4735
$\checkmark$	$\checkmark$	$\checkmark$		30.8831/0.8590	23.8901/0.5617	19.4576
$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	31.5219 / 0.8752	24.0965 / 0.6076	16.8334

TABLE IV: Component Ablation Study.

Low-freq.	High-freq.	$\bm{\mathcal{L}_{\mathrm{freq}}}$	Mean $\uparrow$	$p$ $\uparrow$	$\theta$ $\downarrow$
$\checkmark$			30.7282/0.8552	23.4753/0.5693	18.0932
	$\checkmark$		30.5737/0.8565	22.9142/0.5677	18.2840
$\checkmark$	$\checkmark$		30.8807/0.8603	23.4468/0.5800	17.3958
$\checkmark$	$\checkmark$	$\checkmark$	31.5219 / 0.8752	24.0965 / 0.6076	16.8334

TABLE V: Analysis of Dual-Domain Frequency Interaction.

Image	Stokes	Mean $\uparrow$	$p$ $\uparrow$	$\theta$ $\downarrow$
$\checkmark$		31.2725/0.8665	23.4205/0.5800	17.5793
	$\checkmark$	30.6966/0.8563	23.4003/0.6306	18.1634
$\checkmark$	$\checkmark$	31.5219 / 0.8752	24.0965 / 0.6076	16.8334

TABLE VI: Analysis of Degradation-Conditioned Modulation.

$\bm{\mathcal{L}_{\mathrm{cons}}}$	$\bm{\mathcal{L}_{\mathrm{aop}}^{w}}$	$\bm{\mathcal{L}_{\mathrm{dop}}}$	Mean $\uparrow$	$p$ $\uparrow$	$\theta$ $\downarrow$
$\checkmark$			30.9934/0.8677	23.9921/0.5799	19.1311
$\checkmark$	$\checkmark$		31.3422/0.8701	23.8743/0.5701	17.3279
$\checkmark$		$\checkmark$	31.2095/0.8691	23.9910/0.5935	18.5869
$\checkmark$	$\checkmark$	$\checkmark$	31.5219 / 0.8752	24.0965 / 0.6076	16.8334

TABLE VII: Analysis of Physical Consistency Learning.

Fig. 5: Quantitative analysis of Image–Stokes physical consistency.

Fig. 6: Qualitative comparisons under normal, low-light, and haze conditions. For each scene, we show restored total intensity $s_{0}$ , DoLP $p$ , and AoLP $\theta$ .

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Task	Train	Test	Total
CPDNet [ 7 ]	Normal	85	20	105
PIDSR [ 16 ]	Normal	108	30	138
Sparse-PDM [ 10 ]	Normal	126	20	146
Sparse-PDM [ 10 ]	Haze	69	14	83
Tokyo [ 35 ]	Normal	40	–	40
PLIE [ 11 ]	Low-light	100	30	130

Appendix

TABLE VIII: Dataset composition and degradation condition.

Fig. 8: Sample examples from the constructed dataset.

Fig. 9: Image–Stokes inconsistency maps under normal, low-light, and haze conditions. Lower values indicate better physical consistency.

$\lambda_{1}$	Mean $\uparrow$	$s_{0}$ $\uparrow$	$p$ $\uparrow$	$\theta$ $\downarrow$
0.1	30.962/0.8655	31.874/0.8924	23.974/0.5970	17.462
0.2	31.522/0.8752	32.446/0.9006	24.097/0.6076	16.833
0.3	30.975/0.8651	32.102/0.8926	22.883/0.5841	16.796
0.4	31.109/0.8663	32.071/0.8937	23.079/0.5851	17.295

Appendix

TABLE IX: Effect of the Stokes consistency loss

\mathcal{L}_{\mathrm{cons}}

weight

\lambda_{1}

$\lambda_{2}$	Mean $\uparrow$	$s_{0}$ $\uparrow$	$p$ $\uparrow$	$\theta$ $\downarrow$
0.01	30.665/0.8563	31.601/0.8851	24.075/0.5689	17.245
0.02	30.873/0.8615	31.799/0.8900	23.693/0.5732	17.025
0.03	31.522/0.8752	32.446/0.9006	24.097/0.6076	16.833
0.04	31.175/0.8581	32.005/0.8946	23.888/0.5942	16.657
0.05	30.450/0.8552	31.386/0.8833	22.868/0.5714	16.290

Appendix

TABLE X: Effect of the AoLP loss

\mathcal{L}_{\mathrm{aop}}^{w}

weight

\lambda_{2}

$\tau_{l}$	$\tau_{h}$	All $\downarrow$	$p>0.05$ $\downarrow$	$p>0.10$ $\downarrow$	$p>0.20$ $\downarrow$
0.05	0.15	18.036	13.900	13.950	12.298
0.03	0.20	16.936	14.438	13.320	11.605
0.05	0.20	16.833	13.569	12.701	11.119
0.07	0.20	17.566	14.530	13.534	11.978
0.05	0.25	17.928	14.889	13.880	12.195

Appendix

TABLE XI: Effect of DoLP confidence thresholds

(\tau_{l},\tau_{h})

Fig. 10: Effect of the DoLP feasibility loss $\mathcal{L}_{\mathrm{dop}}$ weight $\lambda_{3}$ .

Fig. 11: Effect of modulation residual-scale initialization $\alpha$ on mean-image PSNR. $\Delta$ PSNR is measured relative to $\alpha=0$ under normal, low-light, haze, and average settings.

Image Restoration

Image restoration uses computational methods to recover useful visual content from noisy, blurred, low-resolution, or otherwise degraded images.

See all 357 papers

Learning Polarization Image Restoration with General Restoration Priors
Sep 25, 2026Chenggong Li, Jinhao Liu, Caiyun Wu +3Image Restoration Polarimetric Imaging
School of Automation Central South University · School of Engineering Westlake University
Architectural Unification for Polarimetric Imaging Across Multiple Degradations
Mar 6, 2026Chu Zhou, Yufei Han, Junda Liao +6Demosaicing Polarimetric Imaging
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.
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

Sep 25, 2026J/K move · Enter open · S save

Learning Polarization Image Restoration with General Restoration Priors

Chenggong Li, Jinhao Liu, Caiyun Wu, Yidong Luo+2

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

Image Restoration Polarimetric Imaging All-in-One Image Restoration

Open paper AskPDF arXiv

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.

Abstract

Image Restoration

Explore similar work

Learning Polarization Image Restoration with General Restoration Priors

Architectural Unification for Polarimetric Imaging Across Multiple Degradations

PolUR: Polarization Aware Multimodal Degradation Perception for Underwater Image Restoration