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.
Figures & tables
Fig. 1: Comparison between proposed PolUR and previous approaches to underwater image restoration. (a) RGB-based UIR methods rely primarily on degraded intensity observations, with degradation priors inferred from the same information-limited RGB inputs, which restricts reliable degradation perception under complex underwater conditions. (b) Polarization-based UIR methods exploit physical cues related to light propagation and scattering, but incorporating these cues alone does not ensure adequate characterization of water conditions and compound degradation states. (c) PolUR unifies physical observation, degradation perception, and restoration decision-making within a framework, jointly using RGB and polarization cues to reason about coexisting degradations and guide restoration.
Fig. 3: Overview of the proposed PolUR framework. (a) and (b) jointly constitute the inference pipeline. Given an RGB image together with P , PolUR produces both an explicit degradation response and a restored image. Specifically, (a) represents UR 2 Net, which performs response-guided image restoration, while (b) denotes the proposed PolarDU for multimodal degradation understanding. (c) illustrates the overall training procedure of PolUR. (d) presents the RPA module embedded in UR 2 Net, which aligns and adaptively integrates the polarization and RGB latent representations.
Fig. 4: The distribution of the Mixpolar.
Fig. 5: Similarity between degraded polarization representations and paired clean references on PURB. Panels (a) and (b) show PSNR and SSIM at an acquisition distance of 45 cm; panels (c) and (d) show the results at 60 cm. Curves represent normal, red-cast, green-cast, and blue-cast conditions. Shaded regions indicate pointwise 95% confidence intervals obtained through scene-level bootstrap resampling.
Fig. 6: Variation in polarization characteristics with scattering severity
Fig. 7: Qualitative comparison of degradation identification using RGB-only and RGB-Polar inputs. Green and red text indicate correct and incorrect attribute predictions.
Fig. 8: Visual comparisons on the MixPolar indoor subset and RGBP. Input and reference images (GT) are shown in the first and last columns, respectively.
Model
Input
Accuracy (%)
F1 (%)
Score
Qwen3-VL-8B
RGB-only
4.40
26.01
15.20
RGB-Polar
1.70
26.96
14.33
InternVL3.5-8B
RGB-only
1.31
32.62
16.97
RGB-Polar
7.33
37.69
22.51
MiniCPM-V 4.5
RGB-only
6.56
30.28
18.42
RGB-Polar
6.02
24.45
15.23
TABLE I: Performance Comparison of Different Methods
Method
MixPolar Indoor
RGBP
PSNR ↑
SSIM ↑
LPIPS ↓
DISTS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
DISTS ↓
UIRPIOM
6.6200
0.3030
0.7930
0.6570
9.3662
0.2467
0.6636
0.4265
UPI2Diff
18.9040
0.4720
0.4236
0.3550
17.0288
0.3937
0.7035
0.4924
UPGD
14.2305
0.5081
0.3387
0.2879
15.8287
0.5195
0.4854
0.4589
SS-UIE
16.8250
0.4300
0.5101
0.3910
16.9519
0.1958
0.4805
0.3270
SLURPP
20.1621
0.7340
0.1760
0.1620
19.4897
0.6037
0.2214
0.1800
TABLE II: Quantitative comparison on the MixPolar indoor and RGBP. The best and second-best distinct values are marked in bold and underlined, respectively, tied best results are both marked in bold. ↑ and ↓ indicate that higher and lower values are better, respectively.
Method
UIQM ↑
URanker ↑
UCIQE ↑
UIRPIOM
0.6370
-0.7039
0.3187
UPI2Diff
2.9569
-0.0010
0.3451
UPGD
1.8900
0.7613
0.2763
SS-UIE
2.2645
0.7897
0.3108
SLURPP
2.5210
0.3099
0.2473
DIE-UIE
2.7413
0.7794
0.3472
TABLE III: Quantitative comparison on MixPolar Outdoor. The best and second-best results are marked in bold and underlined, respectively.
Fig. 9: Visual comparisons on real-world outdoor scenes from MixPolar under direct evaluation without target-domain retraining or fine-tuning. The first and last columns show the input images and PolUR results, respectively.
Fig. 10: Visual comparison of downstream segmentation results. From left to right: raw images, PolUR-restored images, ground-truth masks (GT), and predictions from SegFormer trained and evaluated on raw and restored images, respectively.
National Engineering Research Center for Multimedia Software, School of Computer Science, Wuhan University · School of Computer and Artificial Intelligence, Zhengzhou University · Aerospace Information Research Institute, Henan Academy of Sciences +2