Research map / Image generation / Image restoration
NB-GTR: Narrow-Band Guided Turbulence Removal
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 NB-GTR, which removes atmospheric turbulence from an RGB image guided by a simultaneously captured narrow-band (680 nm) image, using a dual turbulence encoder, re-turb module, and turbulence cue conjunction in a Transformer-based network.
Problem
Single-frame atmospheric turbulence removal performs poorly in diverse real scenes, while multi-frame methods require many captured frames; the paper asks whether a single extra narrow-band image can provide turbulence cues.
Approach
Exploits the wavelength dependence of turbulence: a long-wave narrow-band image has lower tilt variance, and the scaling relation between RGB and narrow-band observations lets the network first induce the turbulence field from their difference and then use it to guide restoration.
Key results
- NB-GTR reaches 25.00 PSNR and 0.8199 SSIM on synthetic data, versus 23.18 / 0.7498 for the best dual-input baseline TurbNet.
- Ablation: removing the narrow-band encoder (W/o narrow) drops PSNR to 24.97 from 25.00 and the full model is best on both metrics.
- Real-data results are qualitative only, using two Sony cameras with a 680 nm, 40 nm bandwidth filter and heat-stove turbulence.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add a narrow-band image to the RGB input | N-IDR · Synthetic data | PSNR: 20.77 -> 20.72 | Hurt |
| Add a narrow-band image to the RGB input | N-IDR · Synthetic data | SSIM: 0.6265 -> 0.6231 | Hurt |
| Add a narrow-band image to the RGB input | Burstormer · Synthetic data | PSNR: 20.10 -> 21.37 | Helped |
| Add a narrow-band image to the RGB input | AT-Net · Synthetic data | SSIM: 0.6008 -> 0.6877 | Helped |
| Add a narrow-band image to the RGB input | FocalNet · Synthetic data | PSNR: 22.44 -> 22.80 | Helped |
| Add a narrow-band image to the RGB input | TurbNet · Synthetic data | SSIM: 0.7167 -> 0.7498 | Helped |
| Use NB-GTR instead of TurbNet (Dual) | Synthetic-data turbulence removal · Synthetic data | PSNR: 23.18 -> 25.00 | Helped |
| Use NB-GTR instead of TurbNet (Dual) | Synthetic-data turbulence removal · Synthetic data | SSIM: 0.7498 -> 0.8199 | Helped |
| Add the re-turb module (ablation) | NB-GTR · Synthetic data | PSNR: 24.89 -> 25.00 | Helped |
| Add the re-turb module (ablation) | NB-GTR · Synthetic data | SSIM: 0.8141 -> 0.8199 | Helped |
| Add the encoder and re-turb module (ablation) | NB-GTR · Synthetic data | PSNR: 24.87 -> 25.00 | Helped |
| Add the encoder and re-turb module (ablation) | NB-GTR · Synthetic data | SSIM: 0.8142 -> 0.8199 | Helped |
Surprising findings
- Retraining baselines with the narrow-band image as extra input gives only modest gains (e.g., TurbNet 22.41 to 23.18 PSNR, N-IDR slightly drops from 20.77 to 20.72).
Limitations
- Requires an additional narrow-band filter, adding acquisition cost.
- Parallel-camera capture is not truly simultaneous, unlike beam-splitter setups.
- Real-data evaluation is only qualitative, with limited scene types.
Experimental setup
Datasets: Places
Benchmarks: synthetic turbulence test set (Places-based) real captured indoor objects and postcards
Metrics: PSNR SSIM
Baselines: N-IDR AT-Net TurbNet Burstormer FocalNet
Builds on: TurbNet phase-to-space simulator spatially adaptive modulation CBAM
Closest papers by meaning
- Turb-Seg-Res: A Segment-then-Restore Pipeline for Dynamic Videos with Atmospheric Turbulence (CVPR 2024 · similarity 0.85)
- Spatio-Temporal Turbulence Mitigation: A Translational Perspective (CVPR 2024 · similarity 0.85)
- Learning Phase Distortion with Selective State Space Models for Video Turbulence Mitigation (CVPR 2025 · similarity 0.84)
- PlaNet: Learning to Mitigate Atmospheric Turbulence in Planetary Images (AAAI 2025 · similarity 0.84)
- Decoupling Scattering: Pseudo-Label Guided NeRF for Scenes with Scattering Media (AAAI 2025 · similarity 0.83)
- PRISM: Rethinking Atmospheric Scattering Reconstruction as a Unified Understanding and Restoration Model for Real-world Dehazing (NeurIPS 2026 · 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.