Research map / Image generation / Image restoration
Unsupervised Pan-Sharpening via Mutually Guided Detail Restoration
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 PAN-MGDR, an unsupervised pan-sharpening method that treats fusion as blind deblurring with a CNN-estimated kernel and mutually guided detail restoration between MS and PAN images.
Problem
Pan-sharpening without ground-truth HRMS images; existing unsupervised methods rely on fixed degradation operators and assumed high-frequency consistency, which are fragile across satellite datasets.
Approach
Model pan-sharpening as blind image deblurring where a CNN (KENet) estimates the LRMS blur kernel under non-negativity and sum-to-one constraints, and a PAN deblurring network uses the fused MS image as guidance to restore the blurred PAN image, forcing the fused image to inherit PAN spatial details.
Key results
- On reduced-resolution GaoFen2, PAN-MGDR reaches PSNR 35.71 and SAM 1.428, versus 31.83 PSNR for the best traditional baseline (IHS).
- On full-resolution WorldView-3, PAN-MGDR achieves HQNR 0.905 vs. 0.849 for GS (best baseline in that column).
- On full-resolution GaoFen2, D_lambda 0.028 and HQNR 0.953, substantially better than PGMAN's 0.861 HQNR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PAN-MGDR instead of the compared methods | Pan-sharpening · GaoFen2, QuickBird, and WorldView-3 reduced-resolution | PSNR: 30.99 -> 35.71; 30.72 -> 31.49; 29.16 -> 31.03 | Helped |
| Use PAN-MGDR instead of the compared methods | Pan-sharpening · GaoFen2, QuickBird, and WorldView-3 full-resolution | HQNR: 0.875 -> 0.953; 0.832 -> 0.862; 0.849 -> 0.905 | Helped |
| Set λ to 1 rather than 3 (ablation) | PAN-MGDR · GaoFen2 reduced-resolution | PSNR: 35.71 -> 32.08 | Hurt |
| Set λ to 5 rather than 3 (ablation) | PAN-MGDR · GaoFen2 reduced-resolution | PSNR: 35.71 -> 35.52 | Hurt |
| Set λ to 7 rather than 3 (ablation) | PAN-MGDR · GaoFen2 reduced-resolution | PSNR: 35.71 -> 35.38 | Hurt |
Surprising findings
- The gradient consistency loss is said to be needed to prevent collapsed solutions even though gradients of HRMS and PAN images may not align.
Limitations
- Gradient consistency loss may compromise fused-image quality since HRMS and PAN gradients need not align.
- Full-resolution evaluation uses the same data as training, and no-reference metrics are used there.
Experimental setup
Datasets: GaoFen2 QuickBird WorldView-3
Benchmarks: GaoFen2 QuickBird WorldView-3
Metrics: PSNR SAM ERGAS Q4 Q8 D_lambda D_S HQNR
Baselines: Brovey IHS GS GFPCA PanGAN PGMAN LDP-Net UCGAN
Builds on: Z-PNN UCGAN
Closest papers by meaning
- SSUN-Net: Spatial-Spectral Prior-Aware Unfolding Network for Pan-Sharpening (AAAI 2025 · similarity 0.90)
- Progressive High-Frequency Reconstruction for Pan-Sharpening with Implicit Neural Representation (AAAI 2024 · similarity 0.89)
- U-Know-DiffPAN: An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening (CVPR 2025 · similarity 0.88)
- Accelerated Diffusion via High-Low Frequency Decomposition for Pan-Sharpening (AAAI 2025 · similarity 0.87)
- Pan-LUT: Efficient Pan-sharpening via Learnable Look-Up Tables (NeurIPS 2025 · similarity 0.87)
- Binarized Neural Network for Multi-spectral Image Fusion (CVPR 2025 · similarity 0.86)
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.