Research map / Image generation / Super-resolution
PIDSR: Complementary Polarized Image Demosaicing and Super-Resolution
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 PIDSR, a joint recurrent two-stage network that performs polarized image demosaicing and super-resolution from a CPFA raw image, with Stokes-aided feature injection to improve DoP and AoP accuracy.
Problem
Polarization camera CPFA raw images need demosaicing and resolution enhancement, but sequential demosaicing then super-resolution amplifies artifacts in degree and angle of polarization (DoP, AoP).
Approach
Formulate joint demosaicing and super-resolution as a series of polarized pixel reconstruction sub-problems solved by a two-stage recurrent pipeline (spatial-physical coherence reconstructor and polarization-aware resolution enhancer), with Stokes-domain features injected via SFI blocks to preserve polarization properties.
Key results
- On synthetic data, PIDSR demosaicing reaches 40.24 dB PSNR on S0 and 33.33 dB on p, outperforming TCPDNet (38.65 dB and 32.26 dB).
- At 4x super-resolution, PIDSR attains 36.31 dB PSNR on S0 versus 36.13 for PSRNet and 31.16 for CPSRNet, with lower AoP MAE than PSRNet (13.05 vs 13.78).
- A polarization-based reflection removal application shows PIDSR→PRR yields fewer artifacts than PID→PISR→PRR and PID→PRR→SISR pipelines (qualitative).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use complementary joint demosaicing and super-resolution rather than sequential stages (ablation) | PIDSR · Synthetic data, demosaicing | MAE (θ): 19.2907 -> 12.2383 | Helped |
| Use complementary joint demosaicing and super-resolution rather than sequential stages (ablation) | PIDSR · Synthetic data, 2× super-resolution | MAE (θHR): 21.8965 -> 12.3520 | Helped |
| Use a two-stage pipeline instead of a single-stage pipeline (ablation) | PIDSR · Synthetic data, demosaicing | MAE (θ): 38.2174 -> 12.2383 | Helped |
| Use a two-stage pipeline instead of a single-stage pipeline (ablation) | PIDSR · Synthetic data, 2× super-resolution | MAE (θHR): 38.2203 -> 12.3520 | Helped |
| Add Stokes feature injection blocks instead of original convolution blocks (ablation) | PIDSR · Synthetic data, demosaicing | MAE (θ): 13.1242 -> 12.2383 | Helped |
| Add Stokes feature injection blocks instead of original convolution blocks (ablation) | PIDSR · Synthetic data, 2× super-resolution | MAE (θHR): 14.1309 -> 12.3520 | Helped |
| Replace PIDSR super-resolution with PSRNet after PIDSR demosaicing (ablation) | PIDSR · Synthetic data, 2× super-resolution | MAE (θHR): 12.5958 -> 12.3520 | Hurt |
| Replace PIDSR demosaicing with TCPDNet before PIDSR super-resolution (ablation) | PIDSR · Synthetic data, 2× super-resolution | MAE (θHR): 13.1530 -> 12.3520 | Hurt |
Surprising findings
- Error rates of DoP and AoP from demosaicing results are much larger than those of S0, and lower spatial resolution correlates with stronger demosaicing artifacts.
Limitations
- Designed for single CPFA raw images, so unsuitable for polarized video reconstruction.
- Cannot handle ordinary CFA raw images because it requires Stokes parameters as input.
- Synthetic dataset is used for most evaluation since public datasets are insufficient.
Experimental setup
Datasets: synthetic polarized dataset rendered with Mitsuba 3
Benchmarks: synthetic polarization test set real polarization camera data
Metrics: PSNR SSIM MAE
Baselines: Polanalyser IGRI2 TCPDNet PSRNet CPSRNet
Builds on: Polanalyser PSRNet CPSRNet TCPDNet Restormer
Closest papers by meaning
- Glossy Object Reconstruction with Cost-effective Polarized Acquisition (CVPR 2025 · similarity 0.85)
- RGB-to-Polarization Estimation: A New Task and Benchmark Study (NeurIPS 2025 · similarity 0.84)
- Quality-Improved and Property-Preserved Polarimetric Imaging via Complementarily Fusing (NeurIPS 2024 · similarity 0.84)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.83)
- PolarFree: Polarization-based Reflection-Free Imaging (CVPR 2025 · similarity 0.82)
- Polarized Color Screen Matting (CVPR 2025 · 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.