Research map / Image generation / Image restoration
Deep Linear Array Pushbroom Image Restoration: A Degradation Pipeline and Jitter-Aware Restoration Network
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 JARNet, a two-stage jitter-aware network (optical flow correction pre-warping plus NAFNet-style SFRes enhancement with coordinate attention and frequency branch) and a CDSM-based pipeline to synthesize degraded linear array pushbroom images from DOTA-v1.0.
Problem
Restoring linear array pushbroom (LAP) remote sensing images, which suffer from distortion and blur caused by camera jitter, with no customized dataset and limited prior methods.
Approach
Combine a CDSM-based finer-grained jitter simulation pipeline with a two-stage restoration: first warp the degraded image by a refined jitter-derived optical flow (OFC block), then enhance with a NAFNet-based network whose SFRes blocks add coordinate attention (spatial) and an FFT frequency branch.
Key results
- JARNet reaches 39.02 dB PSNR and 0.9493 SSIM on the synthetic LAP dataset, 1.28 dB above the best baseline Stripformer (37.74 dB).
- Adding the OFC pre-correction block alone improves PSNR by 0.39 dB over the baseline (38.56 vs 38.17 dB); CoA alone adds 0.33 dB.
- Without CDSM, JARNet drops only slightly (38.50 dB) while pre-corrected baselines drop sharply, supporting CDSM's role in smoothing noisy jitter curves.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use JARNet instead of Stripformer | LAP image restoration · LAP dataset | PSNR (dB): 37.74 -> 39.02; + 1.28dB | Helped |
| Use JARNet instead of Stripformer | LAP image restoration · LAP dataset | GMSD: 0.0326 -> 0.0239; 0.0087 improvements | Helped |
| Add pre-correction to Stripformer (ablation) | Stripformer · LAP dataset | PSNR (dB): 37.74 -> 38.72 | Helped |
| Add CoA (ablation) | JARNet baseline without Freq, CoA, or OFC · LAP dataset | PSNR: 38.17 -> 38.50; 0.33dB | Helped |
| Add frequency branch only (ablation) | JARNet baseline without Freq, CoA, or OFC · LAP dataset | PSNR: 38.17 -> 37.98; 0.19dB | Hurt |
| Add OFC (ablation) | JARNet baseline without Freq, CoA, or OFC · LAP dataset | PSNR: 38.17 -> 38.56; 0.39dB | Helped |
| Add frequency branch and CoA together (ablation) | JARNet · LAP dataset | PSNR: 38.50 -> 38.69 | Helped |
| Remove CDSM from the degradation pipeline (ablation) | JARNet · LAP dataset | PSNR (dB): 39.02 -> 38.50 | Hurt |
| Remove CDSM from the degradation pipeline (ablation) | Stripformer* · LAP dataset | PSNR (dB): 38.72 -> 37.52 | Hurt |
| Use JARNet instead of Stripformer* | Restoration on AID · AID | PSNR (dB): 33.81 -> 34.12 | Helped |
| Use JARNet instead of Stripformer* | Restoration on PatternNet · PatternNet | PSNR (dB): 32.39 -> 33.40 | Helped |
Surprising findings
- Adding the frequency branch alone decreased PSNR by 0.19 dB relative to the baseline (37.98 vs 38.17 dB).
- Without CDSM, pre-correction makes several enhanced baselines worse than their non-enhanced versions.
Limitations
- Evaluated only on a synthetic dataset generated by the authors' own degradation model; real-world results are qualitative only on Yaogan-26 images.
- Ideal jitter model simulates only four sinusoidal components with assumed parameters, which may not capture real jitter complexity.
- Frequency branch alone hurt performance in ablation, and SSIM is not best for JARNet against Stripformer* in the enhanced comparison.
Experimental setup
Datasets: DOTA-v1.0 AID PatternNet
Benchmarks: synthetic LAP dataset (DOTA-v1.0-derived) AID-derived LAP dataset PatternNet-derived LAP dataset
Metrics: PSNR SSIM GMSD GMACs
Baselines: MIMO-UNet+ HINet Uformer-B NAFNet FNAFNet Restormer Stripformer MST MST++ CST DAUHST HDNet
Builds on: NAFNet Coordinate Attention SimpleGate Stripformer
Closest papers by meaning
- Posterior-Mean Rectified Flow: Towards Minimum MSE Photo-Realistic Image Restoration (ICLR 2025 · similarity 0.84)
- From Zero to Detail: Deconstructing Ultra-High-Definition Image Restoration from Progressive Spectral Perspective (CVPR 2025 · similarity 0.83)
- MaIR: A Locality- and Continuity-Preserving Mamba for Image Restoration (CVPR 2025 · similarity 0.83)
- Continuous Optical Zooming: A Benchmark for Arbitrary-Scale Image Super-Resolution in Real World (CVPR 2024 · similarity 0.83)
- BUFF: Bayesian Uncertainty Guided Diffusion Probabilistic Model for Single Image Super-Resolution (AAAI 2025 · similarity 0.82)
- DiffSCI: Zero-Shot Snapshot Compressive Imaging via Iterative Spectral Diffusion Model (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.