Abstract
Objective Underwater optical imaging is severely degraded by absorption and multiple scattering in turbid water, resulting in low contrast, blurred details, and loss of structural information. Polarimetric imaging introduces polarization as an additional physical dimension, enabling partial suppression of backscattered light and enhancement of target visibility. However, the practical performance of underwater polarization image restoration methods is constrained by the scarcity of large-scale, high-quality labeled datasets. Real underwater polarization data are expensive to acquire, difficult to control in terms of scattering parameters, and highly dependent on environmental conditions. Moreover, most existing underwater image simulation and data synthesis methods focus primarily on intensity images and neglect the physical degradation mechanisms of polarization states in scattering media. To address these limitations, this work aims to develop a physics-driven simulation framework for underwater polarization imaging and a dedicated deep learning restoration network that can effectively exploit both spatial and frequency-domain features while preserving polarization consistency. Methods A Mie-scattering-based underwater polarization image simulation model is proposed to generate polarized image datasets under controllable scattering conditions. Based on the Jaffe–McGlamery imaging model, the total detected light is decomposed into target-reflected signal, forward scattering, and backscattering components. Polarization states are represented using the Stokes vector formulation while neglecting circular polarization. The proposed simulation method models the depolarization process induced by multiple scattering events in turbid water. Specifically, the Mueller matrix corresponding to single-particle Mie scattering is calculated using particle diameter, relative refractive index, and scattering angle. By iteratively applying simplified Mueller sub-matrices to the Stokes vectors of target and background light, the cumulative effects of multiple scattering on polarization states are approximated. Forward and backward scattering participation ratios are introduced to account for the probabilistic nature of scattering events. Depth-dependent attenuation is incorporated using relative depth maps obtained from a monocular depth estimation network, which modulate the propagation distance in the simulation process. This framework enables the generation of multi-distance and multi-concentration underwater polarization images with controllable physical parameters. Based on the simulated dataset and real underwater data, a multiscale spatial–frequency feature restoration network (MSFNet) is designed for underwater polarization image restoration. The network takes four-channel polarization images acquired by a division-of-focal-plane (DoFP) polarimetric camera as input. A symmetric encoder–decoder architecture is adopted, with wavelet feature fusion blocks serving as the core modules. Wavelet convolution layers decompose feature maps into low- and high-frequency subbands, allowing the network to explicitly capture structural information and fine textures while suppressing scattering-induced blur. A multiscale feature fusion module with convolutional block attention mechanisms is employed to aggregate contextual information across different spatial scales. Additionally, a shallow feature extraction module is introduced to stabilize low-level representations and enhance robustness under severe scattering conditions. To further constrain the restoration process, a composite loss function is constructed, consisting of spatial-domain reconstruction loss, frequency-domain consistency loss, and polarization consistency loss. The polarization loss enforces structural similarity constraints on intensity, degree of linear polarization, and angle of polarization (AoP), ensuring that the restored images remain physically consistent with polarization imaging principles. Results and Discussions Extensive experiments are conducted on both simulated and real underwater polarization datasets. An underwater active polarization imaging experimental system is built using a 532 nm laser source, polarization optics, and a DoFP polarimetric camera to collect real data under different turbidity levels. The proposed MSFNet is trained using a mixture of simulated and real image pairs. Qualitative comparisons demonstrate that the proposed method effectively suppresses scattering artifacts, enhances contrast, and recovers fine structural details under scattering conditions. Compared with conventional contrast enhancement methods and physics-based restoration methods, the proposed approach achieves superior visual quality without introducing noticeable artifacts. Transformer-based underwater restoration networks exhibit limited effectiveness in heavily turbid scenarios, whereas the proposed spatial–frequency joint modeling strategy shows stronger robustness. Quantitative evaluations using peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) further confirm the advantages of the proposed method on both simulated and real datasets. Ablation studies indicate that the wavelet-based frequency decomposition and polarization consistency loss each contribute significantly to performance improvements. Additional experiments reveal that training with simulated data substantially enhances generalization capability and reduces artifacts when restoring real underwater polarization images. Polarization parameter analysis indicates that AoP recovery remains more challenging due to its high sensitivity to noise and scattering-induced randomization. In contrast, the proposed method achieves stable and reliable recovery of DoLP distributions, even under high turbidity conditions. Conclusions This study presents a Mie-scattering-simulation-driven underwater polarization image restoration framework that integrates physically grounded data generation with a multiscale spatial–frequency deep learning network. The proposed simulation model accurately simulates polarization degradation mechanisms in turbid water and provides a controllable, efficient source of training data. The MSFNet effectively combines spatial and frequency-domain representations while enforcing polarization consistency, leading to robust restoration performance on both simulated and real underwater datasets. Experimental results validate the effectiveness and generalization capability of the proposed approach. Future work will focus on improving angle of polarization recovery under extreme scattering conditions and extending the simulation framework to more complex underwater environments and illumination configurations.
| Translated title of the contribution | Mie散射仿真驱动的水下偏振图像多尺度特征复原方法 |
|---|---|
| Original language | English |
| Article number | 1811001 |
| Journal | Guangxue Xuebao/Acta Optica Sinica |
| Volume | 46 |
| Issue number | 18 |
| DOIs | |
| Publication status | Published - 2026 |
| Externally published | Yes |
Keywords
- deep learning
- image restoration
- mie scattering
- underwater polarization imaging
Fingerprint
Dive into the research topics of 'Mie-Scattering-Simulation-Driven Multiscale Feature Restoration Method for Underwater Polarization Images'. Together they form a unique fingerprint.Cite this
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver