Research map / Image generation / Image restoration
Quality-Improved and Property-Preserved Polarimetric Imaging via Complementarily Fusing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a three-phase neural fusing framework that combines a noisy short-exposure and a blurry long-exposure polarized snapshot to produce a clean, clear snapshot with accurate DoP and AoP.
Problem
Polarimetric imaging suffers from a trade-off: short exposures give noisy images that degrade DoP and AoP, while long exposures cause motion blur that over-smooths them. The goal is to fuse a noisy-blurry pair while preserving polarization properties.
Approach
Three phases tailored to polarization: irradiance restoration of S0 guided by color (S0 of blurry input) and structure (S1,S2 of noisy input) cues via CSCF modules; polarization reconstruction of DoP/AoP in a Cartesian (x,y) representation using CAG coherence volumes and CI cross-attention Transformer modules; and image-domain artifact suppression with a U-Net and a polarization consistency loss.
Key results
- On synthetic data the framework reaches PSNR-p 29.23 and PSNR-S0 39.05, outperforming PLIE+ (27.98 / 39.01) and PolDeblur+ (25.31 / 39.04).
- Replacing the Cartesian DoP/AoP reconstruction with direct reconstruction drops PSNR-p from 29.23 to 28.04.
- Fusing noisy and blurry inputs improves over single-input PLIE and PolDeblur, supporting the complementary-fusion premise.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the three-phase complementarily fusing framework instead of PLIE+ | Polarimetric imaging · Synthetic test set | PSNR-p: 27.98 -> 29.23 | Helped |
| Use the three-phase complementarily fusing framework instead of PLIE+ | Polarimetric imaging · Synthetic test set | PSNR-θ: 16.93 -> 16.96 | Helped |
| Replace the irradiance restoration phase with LSD2 (ablation) | Ours · Synthetic test set | PSNR-p: 29.23 -> 29.12 | Hurt |
| Add CSCF modules instead of vanilla convolution layers (ablation) | Ours · Synthetic test set | PSNR-p: 29.21 -> 29.23 | Helped |
| Add the CAG module (ablation) | Ours · Synthetic test set | PSNR-p: 29.07 -> 29.23 | Helped |
| Add the CI modules (ablation) | Ours · Synthetic test set | PSNR-p: 28.92 -> 29.23 | Helped |
| Reconstruct DoP and AoP in Cartesian coordinates rather than directly (ablation) | Ours · Synthetic test set | PSNR-p: 28.04 -> 29.23 | Helped |
| Add the artifact suppression refinement phase (ablation) | Ours · Synthetic test set | PSNR-p: 29.02 -> 29.23 | Helped |
Surprising findings
- Existing conventional-image fusion methods (SelfIR, D2HNet) perform worse than single-image polarized methods on DoP/AoP, indicating that polarization-unaware fusion is inadequate.
Limitations
- Cannot reconstruct polarized video; only single snapshots.
- Cannot fuse conventional RGB images because the first phase requires Stokes parameters.
- Evaluated only on a synthetic dataset derived from PLIE plus a small set of real captures; no public dataset exists for this setting.
Experimental setup
Datasets: PLIE dataset synthetic noisy-blurry polarized pairs generated from PLIE
Benchmarks: synthetic noisy-blurry polarized dataset (derived from PLIE) real captured noisy-blurry pairs
Metrics: PSNR SSIM
Baselines: PLIE PLIE+ PolDeblur PolDeblur+ LSD2 LSFNet SelfIR D2HNet
Builds on: PLIE PolDeblur CBAM deformable convolution U-Net
Closest papers by meaning
- Glossy Object Reconstruction with Cost-effective Polarized Acquisition (CVPR 2025 · similarity 0.84)
- PIDSR: Complementary Polarized Image Demosaicing and Super-Resolution (CVPR 2025 · similarity 0.84)
- PolarFree: Polarization-based Reflection-Free Imaging (CVPR 2025 · similarity 0.83)
- High-Fidelity Polarimetric Implicit 3D Reconstruction with View-Dependent Physical Representation (AAAI 2025 · similarity 0.83)
- PolarMatte: Fully Computational Ground-Truth-Quality Alpha Matte Extraction for Images and Video using Polarized Screen Matting (CVPR 2024 · similarity 0.83)
- Learning to Remove Wrinkled Transparent Film with Polarized Prior (CVPR 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.