Research map / Video / Video segmentation and tracking
Hazy Low-Quality Satellite Video Restoration Via Learning Optimal Joint Degradation Patterns and Continuous-Scale Super-Resolution Reconstruction
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 ODPNet, an end-to-end network that jointly dehazes and performs continuous-scale super-resolution on hazy low-quality satellite videos, using feature-level soft dehazing, spatiotemporal self-attention, a tri-branch cross-aggregation block and a cross-scale Top-k selection Transformer.
Problem
Hazy low-resolution satellite videos suffer double degradation (haze plus downsampling) with weak textures and complex motion; existing two-stage dehazing-then-VSR pipelines accumulate errors and most satellite VSR methods only support integer scales.
Approach
Models haze and downsampling as a combined degradation, removes haze at the feature level by predicting atmospheric-scattering-inspired factors K and B, fuses spatiotemporal features with joint spatial/temporal self-attention, and uses Top-k selection over multi-scale latent codes for arbitrary-scale upsampling.
Key results
- ODPNet reaches 33.65 PSNR at x2.1 and 31.84 at x3.2 on Clip001 (haze 0.7,0.35 / 1.0,0.60 settings), exceeding the best TSP combination in Table 1.
- On x4 across ten clips, ODPNet with 7.3M parameters outperforms six TSP baselines; e.g., Clip005 reaches 37.28 PSNR vs at most 26.47 for the best baseline shown.
- Ablation: PFSD gives +1.91 dB over w/o PFSD at x3.2 on Clip001; SSA with both temporal and spatial attention gains +0.77 dB over no self-attention.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ODPNet instead of the strongest listed two-stage baseline | hazy satellite VSR · Clip001-Clip010, ×4 | PSNR | Helped |
| add prior-based feature soft dehazing (PFSD) (ablation) | ODPNet · Clip001, ×2.1 | PSNR: 32.13 -> 33.65 | Helped |
| add prior-based feature soft dehazing (PFSD) (ablation) | ODPNet · Clip001, ×3.2 | PSNR: 29.93 -> 31.84 | Helped |
| add prior-based feature soft dehazing (PFSD) (ablation) | ODPNet · Clip001, ×4 | PSNR: 28.34 -> 28.63 | Helped |
| add spatial self-attention to temporal self-attention (ablation) | ODPNet SSA · Clip001-010, haze (0.7, 0.35), ×4 | PSNR: 28.86 -> 29.30 | Helped |
| add temporal and spatial self-attention (ablation) | ODPNet SSA · Clip001-010, haze (0.7, 0.35), ×4 | PSNR: 28.53 -> 29.30 | Helped |
| add interaction aggregation (IA) (ablation) | ODPNet TCB · Clip001, haze (0.7, 0.35), scales (2.1, 2.7, 3.2, 4) | PSNR: 29.93 -> 30.67 | Helped |
| use Top-k selection rather than all cross-scale features (ablation) | ODPNet TST · Clip001-010, haze (1.0, 0.6), ×2.1 | PSNR: 36.05 -> 37.02 | Helped |
| set Top-k thresholds to [50%, 80%] instead of [50%, 66%] (ablation) | ODPNet TST · Clip001-010, haze (1.0, 0.6), ×2.1 | PSNR: 36.85 -> 37.02 | Helped |
| set Top-k thresholds to [50%, 80%] instead of [50%, 75%] (ablation) | ODPNet TST · Clip001-010, haze (1.0, 0.6), ×2.1 | PSNR: 35.97 -> 37.02 | Helped |
Surprising findings
- Chaining state-of-the-art dehazing and VSR models in two stages (TSP) with larger parameter counts still underperforms the single end-to-end ODPNet at 7.3M parameters.
Limitations
- Evaluation relies on synthetic haze and bicubic degradation; real hazy ISS results are only qualitative.
- Only three test scenes in the main setup and the dataset is not released.
Experimental setup
Datasets: Jilin-1 satellite videos (synthetic haze and bicubic downsampling) real hazy ISS satellite videos (qualitative)
Benchmarks: Jilin-1 satellite video test clips Clip001-Clip010
Metrics: PSNR SSIM
Baselines: TSP-FTM TSP-DEL TSP-MEM TSP-FC TSP-DC TSP-MC TSP-FM TSP-DM TSP-FL TSP-ML FFANet DehazeFormer
Builds on: atmospheric scattering model RTDRAM DehazeFormer LIFF Meta-SR PAConv
Closest papers by meaning
- FMA-Net: Flow-Guided Dynamic Filtering and Iterative Feature Refinement with Multi-Attention for Joint Video Super-Resolution and Deblurring (CVPR 2024 · similarity 0.86)
- Low-Res Leads the Way: Improving Generalization for Super-Resolution by Self-Supervised Learning (CVPR 2024 · similarity 0.84)
- Blur-aware Spatio-temporal Sparse Transformer for Video Deblurring (CVPR 2024 · similarity 0.84)
- Efficient Visual State Space Model for Image Deblurring (CVPR 2025 · similarity 0.84)
- Tokenize Image Patches: Global Context Fusion for Effective Haze Removal in Large Images (CVPR 2025 · similarity 0.84)
- Unsupervised Degradation Representation Aware Transform for Real-World Blind Image Super-Resolution (AAAI 2025 · similarity 0.83)
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.