Research map / Label-efficient and robust learning / Recognition and long-tail
Deep Unfolded Network with Intrinsic Supervision 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 DISPNet, an interpretable deep unfolded network for pan-sharpening that unfolds a variational model with spatial consistency and spectral projection priors via half-quadratic splitting, with intrinsic supervision on intermediate features.
Problem
Deep pan-sharpening methods lack supervision of complementary PAN/MS information in intermediate layers and are black-box, with low interpretability.
Approach
Formulate pan-sharpening as variational minimization with a spatial consistency prior (joint component decomposition of PAN and MS) and a spectral projection prior, unfold the half-quadratic splitting solver into K stages, and supervise intermediate spatial components and spectral responses with an intrinsic supervision loss.
Key results
- On reduced-scale QuickBird, the method reports ERGAS 1.0445, RMSE 2.2080, SAM 1.3870, SSIM 0.9668, outperforming listed baselines such as P2Sharpen (ERGAS 1.3302).
- On full-scale GaoFen-2, QNR reaches 0.9223 versus 0.8574 for the next best (MMNet), with D_lambda 0.0344 and D_s 0.0449.
- Ablation on QuickBird shows that adding intrinsic supervision lowers ERGAS from 1.0865 to 1.0445 and improves SAM and SSIM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the proposed deep unfolded network with intrinsic supervision | Pan-sharpening · GaoFen-2 reduced-scale | ERGAS: 1.3011 -> 1.2490 | Helped |
| use the proposed deep unfolded network with intrinsic supervision | Pan-sharpening · QuickBird reduced-scale | ERGAS: 1.3302 -> 1.0445 | Helped |
| use the proposed deep unfolded network with intrinsic supervision | Pan-sharpening · WorldView-II reduced-scale | ERGAS: 5.3133 -> 5.1130 | Helped |
| use the proposed deep unfolded network with intrinsic supervision | Pan-sharpening · GaoFen-2 full-scale | QNR: 0.8574 -> 0.9223 | Helped |
| use the proposed deep unfolded network with intrinsic supervision | Pan-sharpening · QuickBird full-scale | QNR: 0.9252 -> 0.9592 | Helped |
| use the proposed deep unfolded network with intrinsic supervision | Pan-sharpening · WorldView-II full-scale | QNR: 0.8971 -> 0.9050 | Helped |
| add the spatial consistency prior (ablation) | Ours-IV pan-sharpening model · QuickBird reduced-scale | ERGAS: 1.3766 -> 1.0865 | Helped |
| add the spectral projection prior (ablation) | Ours-IV pan-sharpening model · QuickBird reduced-scale | ERGAS: 1.1195 -> 1.0865 | Helped |
| add intrinsic supervision loss (ablation) | Pan-sharpening model with both priors · QuickBird reduced-scale | ERGAS: 1.0865 -> 1.0445 | Helped |
| increase unfolding stages from K=1 to K=4 (ablation) | Ours pan-sharpening model · QuickBird reduced-scale | ERGAS: 1.0836 -> 1.0445 | Helped |
| increase unfolding stages from K=4 to K=6 (ablation) | Ours pan-sharpening model · QuickBird reduced-scale | ERGAS: 1.0445 -> 1.0355 | Helped |
Surprising findings
- Full-scale non-reference metrics (D_lambda, D_s) do not always align with visual perception, attributed to limited spectral and spatial reference assumptions.
Limitations
- Full-scale non-reference metrics rely on limited assumptions and can be inconsistent with visual quality.
- Only four unfolding stages chosen for trade-off; more stages increase training and testing time.
Experimental setup
Datasets: GaoFen-2 QuickBird WorldView-II
Benchmarks: GaoFen-2 QuickBird WorldView-II
Metrics: ERGAS RMSE SAM SSIM QNR D_lambda D_s
Baselines: GS IHS SFIM BDSDPC PanNet PanGAN GPPNN MMNet P2Sharpen
Builds on: half-quadratic splitting Wald's protocol PanNet GPPNN
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 0.87)
- Unsupervised Pan-Sharpening via Mutually Guided Detail Restoration (AAAI 2024 · similarity 0.86)
- U-Know-DiffPAN: An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening (CVPR 2025 · similarity 0.84)
- Pan-LUT: Efficient Pan-sharpening via Learnable Look-Up Tables (NeurIPS 2025 · similarity 0.84)
- Understanding Sharpness Dynamics in NN Training with a Minimalist Example: The Effects of Dataset Difficulty, Depth, Stochasticity, and More (ICML 2025 · similarity 0.84)
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.