Research map / Image generation / Image restoration
SSUN-Net: Spatial-Spectral Prior-Aware Unfolding Network for Pan-Sharpening
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 SSPF, a spatial-spectral physical-prior pan-sharpening framework unfolded via Split Bregman into the lightweight deep unfolding network SSUN-Net with multiscale prior structure (MPS).
Problem
Existing deep unfolding networks for pan-sharpening rely on implicit deep priors and single-scale prior features, ignoring physical spatial-spectral priors of MS and PAN and causing spatial and spectral distortions.
Approach
Formulate pan-sharpening as a constrained minimization with explicit spectral (gradient-based) and spatial (PAN-gradient-consistency) priors solved by Split Bregman, unfold it into stages with learnable operators, and add a multiscale prior structure (MSEM/MSAM) with cross-domain interaction.
Key results
- On WV-III simulated data SSUN-Net reaches ERGAS 2.9054, SSIM 0.9292, PSNR 31.116, the best among compared methods.
- On real GaoFen 2 data SSUN-Net achieves QNR 0.8591 versus 0.8454 for the next best (MDCUN) and reduces Dλ to 0.0658.
- SSUN-Net uses 2.6698 GFLOPs and 0.2934M parameters, fewer FLOPs than the other DUNs in Table 1 except LGTEUN-type ones; ablation shows removing MPS priors increases ERGAS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SSUN-Net instead of DISP | pan-sharpening · WV-II simulated | ERGAS: 0.8759 -> 0.8349 | Helped |
| use SSUN-Net instead of DISP | pan-sharpening · WV-III simulated | ERGAS: 3.0096 -> 2.9054 | Helped |
| use SSUN-Net instead of SFINet++ | pan-sharpening · GF-2 simulated | ERGAS: 0.4376 -> 0.4286 | Helped |
| use SSUN-Net instead of DISP | pan-sharpening · GF-2 real | QNR: 0.8437 -> 0.8591 | Helped |
| add CFI rather than replace it with Dense Block (ablation) | SSUN-Net · GF-2 simulated | ERGAS: 0.4563 -> 0.4286 | Helped |
| add ISF rather than replace it with Dense Block (ablation) | SSUN-Net · GF-2 simulated | ERGAS: 0.4407 -> 0.4286 | Helped |
| add attention in MSAM and MSEM rather than replace it with Dense Block (ablation) | SSUN-Net · GF-2 simulated | ERGAS: 0.4330 -> 0.4286 | Helped |
| add spatial prior Ea(H|P) (ablation) | SSUN-Net · GF-2 simulated | ERGAS: 0.4438 -> 0.4286 | Helped |
| add spectral prior Ee(H|L) (ablation) | SSUN-Net · GF-2 simulated | ERGAS: 0.4436 -> 0.4286 | Helped |
| add MPS and both physical priors (ablation) | SSUN-Net · GF-2 simulated | ERGAS: 0.4672 -> 0.4286 | Helped |
| add MPS (ablation) | SSUN-Net · GF-2 simulated | ERGAS: 0.4438 -> 0.4286 | Helped |
| add ISF, CFI, attention, and both physical priors (ablation) | SSUN-Net · GF-2 simulated | ERGAS: 0.4697 -> 0.4286 | Helped |
Surprising findings
- Plugging the MPS structure into GPPNN (Net8) reduces ERGAS only modestly, while the full SSUN-Net without MPS (Net7) is worse, indicating the gains depend on the whole unfolding design.
- Removing physical priors entirely (Net6) still yields reasonable performance, with ERGAS change of about 8% versus the default.
Limitations
- Gains on simulated data are modest in some comparisons (e.g., ERGAS improvements of 2-5% over DISP on some datasets).
- Evaluation limited to three satellite datasets and a single real-data set without high-resolution references.
Experimental setup
Datasets: WorldView II WorldView III GaoFen 2
Benchmarks: WorldView II WorldView III GaoFen 2
Metrics: PSNR SSIM SAM SCC ERGAS Dλ Ds QNR
Baselines: GPPNN DISP NLUNet LGTEUN MDCUN SFINet++ HFIN WINET PDDN INNF
Builds on: GPPNN Split Bregman method Wald protocol
Closest papers by meaning
- Unsupervised Pan-Sharpening via Mutually Guided Detail Restoration (AAAI 2024 · similarity 0.90)
- Deep Unfolded Network with Intrinsic Supervision for Pan-Sharpening (AAAI 2024 · similarity 0.90)
- U-Know-DiffPAN: An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening (CVPR 2025 · similarity 0.88)
- Pan-LUT: Efficient Pan-sharpening via Learnable Look-Up Tables (NeurIPS 2025 · similarity 0.86)
- Accelerated Diffusion via High-Low Frequency Decomposition for Pan-Sharpening (AAAI 2025 · similarity 0.86)
- Detail Matters: Mamba-Inspired Joint Unfolding Network for Snapshot Spectral Compressive Imaging (AAAI 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.