Polarization-Unet: Deep Polarimetric Analysis

• Polarization-Unet is a deep encoder–decoder CNN that utilizes multi-channel polarization data to recover detailed 3D surface structures and cosmological features.
• It adapts a U-Net backbone with physics-guided loss functions and attention mechanisms to fuse complex polarization inputs across various imaging scenarios.
• Empirical results demonstrate significant improvements over traditional baselines, achieving lower mean angular errors and enhanced image fusion quality.

Polarization-Unet constitutes a class of deep encoder–decoder convolutional neural network (CNN) methods that leverage polarization information in imaging and remote sensing contexts. These models adopt adaptations of the U-Net backbone, often with architectural enhancements tailored to the physics and statistics of polarization-resolved data, including but not limited to photometric surface normal estimation, polarization image fusion, and cosmological parameter inference from polarization maps. Central to their operation is the exploitation of polarization-sensitive multi-channel input, with end-to-end training to map directly to physically- or semantically-relevant outputs.

1. Core Methodological Principles

Polarization-Unet models consistently share these features:

• Multi-Channel Polarization Input: Data is represented as multi-channel stacks encoding polarization information. This may be the raw intensity at several polarizer angles $I(\theta)$, Stokes parameters $(S_0, S_1, S_2)$, degree of linear polarization (DOLP), angle of polarization (AOP), or composite representations.
• U-Net-based Encoder–Decoder: The backbone follows the U-Net paradigm: symmetric encoding and decoding paths, with skip connections for local detail preservation.
• Physics Guidance via Loss or Architecture: While mappings are learned end-to-end, either the output targets (e.g., 3D normals or rotation angle $\alpha(\mathbf{n})$) or intermediate representations enforce physical interpretability.
• Joint or Multi-Task Capability: In advanced variants, multiple related quantities are jointly predicted (e.g., normal vector components, cosmological fields).
• Evaluation on Specialized Benchmarks: Performance is compared both qualitatively and quantitatively against traditional physics-driven and deep learning baselines, typically using metrics tailored to the task domain (e.g., mean angular error, structural similarity).

2. Surface Normal Recovery from Polarization Imagery

The Polarization-U-Net for surface normal estimation, as developed by Mortazavi et al. (Mortazavi et al., 2024), leverages passive polarization imaging for high-resolution 3D shape inference:

• Input: Four polarization-filtered images per scene, $I(\theta)$ at $$\theta = \{0^\circ, 45^\circ, 90^\circ, 135^\circ\}$$; raw intensity stack, processed pixelwise.
• Physical Model: Intensity at each pixel follows a sinusoidal model,

$I(\theta) = \frac{I_\max + I_\min}{2} + \frac{I_\max - I_\min}{2} \cos(2(\theta - \phi))$

where $\phi$ relates to the surface normal azimuth.

• Network Design: ResNet-18 encoder; U-Net-style decoder; no explicit physical parameter input, fusing cues via architecture.
• Output: Per-pixel 3D normal vector, normalized postprocessing for unit length.
• Loss: Mean angular error (MAE) between predicted and ground-truth normal vectors:

$$\mathcal{L}_{\mathrm{MAE}} = \frac{1}{N}\sum_{i=1}^N \arccos(\hat{\mathbf{n}}_i \cdot \mathbf{n}_i^{\mathrm{gt}})$$

• Training: Adam optimizer, $10^{-4}$ learning rate, 32 batch size, 100 epochs, random crops as only augmentation.
• Evaluation: On the DeepSfP benchmark, achieves an MAE of 18.06°, outperforming physics-based baselines (41.44°–49.03° MAE) and matching or slightly exceeding prior learning-based approaches (Mortazavi et al., 2024).

This architecture enables recovery of fine-scale surface details and operates robustly under diverse lighting and object material conditions.

3. Polarization Image Fusion in Complex Luminance Environments

For multi-modal fusion of polarization data, the MLS-UNet (Luminance-Aware Multi-Scale UNet), presented by Zeng and collaborators (Huang et al., 28 Oct 2025), integrates advanced feature fusion with explicit luminance guidance:

• Input: $S_0$ (total intensity) and DOLP images; four-direction raw polarized images used for ground-truth derivation.
• Encoder: Double-convolution blocks with CBAM attention; multi-stage Brightness-Branch injects spatial attention coefficients based on normalized luminance $(S_0, S_1, S_2)$0 at each scale to address contrast disparities.
• Global–Local Fusion: At the bottleneck, Swin Transformer-style windowed self-attention captures both contextual and local correlations; residual links and MLP further restructure feature representations.
• Decoder: Brightness-Enhancement module concatenates fused features with luminance information and applies nonlinear correction to adaptively restore correct luminance–texture balance in the reconstructed image.
• Loss: Weighted sum of multiple quality measures—SSIM, $(S_0, S_1, S_2)$1, contrast, texture (gradient), and $(S_0, S_1, S_2)$2-regularization.
• Dataset: MSP comprises 1,000 real-scene polarized image sets across 17 scene types and lighting conditions.
• Performance: Outperforms seven SOTA baselines on public benchmarks (e.g., 8.6% MS-SSIM gain, 60.6% SD gain on MSP test set; similar leads on PIF and GAND datasets) (Huang et al., 28 Oct 2025).

The architecture demonstrates precise fusion and robust feature retention, particularly in shadowed or high-contrast scenarios.

4. Cosmic Polarization Rotation Reconstruction

In CMB analysis, ResUNet-CMB (a Polarization-UNet variant) reconstructs cosmic polarization rotation $(S_0, S_1, S_2)$3 and related fields from Stokes $(S_0, S_1, S_2)$4 maps affected by lensing and reionization (Guzman et al., 2021):

• Input: $(S_0, S_1, S_2)$5 tensors of observed $(S_0, S_1, S_2)$6 maps per sky patch.
• Architecture: Four-level residual U-Net (skip connections across conv pairs; SELU activations; batch-norm in main and residual paths). Decoder mirrors encoder; outputs fork into four heads predicting (1) polarization rotation $(S_0, S_1, S_2)$7, (2) lensing convergence $(S_0, S_1, S_2)$8, (3) patchy reionization $(S_0, S_1, S_2)$9, and (4) primordial $\alpha(\mathbf{n})$0-mode.
• Physical Transformations: Rotation, lensing, and modulation are modeled physically; the network learns to disentangle these effects from mixed map data.
• Loss: Joint mean squared error over all outputs.

$\alpha(\mathbf{n})$1

• Bias Calibration: Harmonic space calibration coefficients $\alpha(\mathbf{n})$2 ensure unbiased reconstruction of power spectra.
• Benchmarking: Variance of the $\alpha(\mathbf{n})$3 reconstruction achieves parity with idealized iterative quadratic estimator methods (for $\alpha(\mathbf{n})$4 in low-noise conditions), and exhibits substantial superiority over standard quadratic estimators on lensed CMB (Guzman et al., 2021).

ResUNet-CMB implicitly learns delensing and demodulation without explicit map preprocessing, extracting physically meaningful anisotropy fields.

5. Quantitative Summary and Comparative Results

Method/Domain	Task	Key Metric & Value	Baseline Comparison
Polarization-U-Net	Surface Normal Estimation	MAE = 18.06° (Mortazavi et al., 2024)	Physics (41.44°–49.03° MAE)
MLS-UNet	Polarization Fusion	MS-SSIM +8.6%, SD +60.6% vs avg baseline	SOTA fusion models
ResUNet-CMB	Cosmological Rotation	Near-iterative QE variance to $\alpha(\mathbf{n})$5	QE, iterative QE methods

Polarization-Unet architectures consistently demonstrate substantial gains in fine-scale detail recovery and physical disentanglement over traditional analytic or shallow-learning baselines across diverse imaging domains.

6. Limitations and Prospective Advances

Despite empirical effectiveness across multiple applications, existing Polarization-Unet approaches face notable constraints:

• Input Device Requirements: Necessitate polarization cameras or precise rotation hardware.
• Generalization Scope: Domain shift to new objects, material types, or cosmological scenarios may impair accuracy without further retraining.
• Explicit Physics Constraints: Learned models may not guarantee physical admissibility (e.g., violate Fresnel law constraints or hallucinate implausible outputs under extreme specularity or noise).
• Data Demands: Require substantial annotated or simulated training datasets tailored to the specific target task.

Ongoing research suggests several extensions: incorporating physical parameter maps (e.g., $\alpha(\mathbf{n})$6, $\alpha(\mathbf{n})$7) as input; multi-task learning of additional physical or material parameters; adaptation to dynamic (“in the wild”) scenes; and integration with explicit model-based pipelines for hybrid inversion (Mortazavi et al., 2024, Huang et al., 28 Oct 2025).

7. Context and Significance

Polarization-Unet and related architectures establish a high-performance, scalable, and physically-motivated framework for extracting structural, material, or cosmological information from polarization-resolved imaging. By marrying deep feature hierarchies and skip connection-based preservation of spatial detail with polarization-specific guidance, these models enable a range of remote and scientific sensing tasks, bypassing many of the limitations of traditional analytical inversion or handcrafted feature pipelines. Their demonstrated superiority in surface reconstruction, image fusion, and CMB analysis marks them as foundational methodologies for next-generation polarization imaging research (Mortazavi et al., 2024, Huang et al., 28 Oct 2025, Guzman et al., 2021).