Research map / Image generation / Image restoration
OTPNet: ODE-inspired Tuning-free Proximal Network for Remote Sensing Image Fusion
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 OTPNet, an unfolded Douglas-Rachford proximal network whose proximal blocks use a fourth-order Runge-Kutta ODE-style update, for tuning-free remote sensing image fusion (pan-sharpening, HSR, MHF).
Problem
Deep learning remote sensing image fusion relies on hand-designed architectures and hyperparameter tuning, with limited interpretability and adaptability.
Approach
Splits fusion into spatial (LR-MS) and spectral (PAN) proximal subproblems solved via Douglas-Rachford splitting, and realizes each proximal operator as a network block based on a fourth-order Runge-Kutta scheme with learnable step-like parameters, avoiding manual tuning.
Key results
- On GaoFen2, OTPNet reaches 49.340 PSNR and 0.992 SSIM for pan-sharpening, the best among the twelve compared methods.
- On WorldView3, OTPNet achieves 39.255 PSNR vs. 38.869 for LAGNet (the strongest baseline in the table), with ERGAS 2.104.
- Ablation: 4th-order RK (RK4) reaches 39.255 PSNR vs. 37.094 for forward Euler (ResNet-style) on WorldView3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Douglas-Rachford splitting instead of GP, ADMM, or HQS (ablation) | OTPNet fusion optimization · WorldView3 | PSNR: 35.833, 38.133, 39.068 -> 39.255 | Helped |
| Use fourth-order Runge-Kutta instead of lower-order numerical schemes (ablation) | OTPNet proximal network · WorldView3 | PSNR: 37.094, 37.297, 38.117, 38.420, 39.050 -> 39.255 | Helped |
| Use tuning-free parameters with fixed architecture (ablation) | OTPNet · WorldView3 | PSNR: 38.859 -> 39.067 | Helped |
| Use tuning-free architecture with fixed parameters (ablation) | OTPNet · WorldView3 | PSNR: 38.859 -> 39.123 | Helped |
| Use tuning-free parameters and tuning-free architecture (ablation) | OTPNet · WorldView3 | PSNR: 38.859 -> 39.255 | Helped |
Limitations
- Evaluated only on remote sensing fusion; generality to other multisource fusion tasks is future work.
- Theoretical error analysis is local truncation error only and does not establish global convergence of the learned network.
- Full hardware and compute details are not stated in the main text.
Experimental setup
Datasets: WorldView-3 GaoFen-2 QuickBird PaviaCentre Botswana4 Chikusei CAVE Harvard NTIRE2020
Benchmarks: WorldView-3 GaoFen-2 QuickBird
Metrics: PSNR SSIM SAM ERGAS
Baselines: Brovey GS IHS SFIM PANNet SRPPNN GPPNN MUCNN MDCUN LAGNet PMACNet PANFormer
Builds on: Douglas-Rachford splitting ResNet Runge-Kutta numerical scheme GPPNN
Closest papers by meaning
- A Selective Re-learning Mechanism for Hyperspectral Fusion Imaging (CVPR 2025 · similarity 0.85)
- Meta-Learning Hyperparameters for Parameter Efficient Fine-Tuning (CVPR 2025 · similarity 0.85)
- Binarized Neural Network for Multi-spectral Image Fusion (CVPR 2025 · similarity 0.84)
- OTIAS: OcTree Implicit Adaptive Sampling for Multispectral and Hyperspectral Image Fusion (AAAI 2025 · similarity 0.84)
- HetSSNet: Spatial-Spectral Heterogeneous Graph Learning Network for Panchromatic and Multispectral Images Fusion (ICML 2025 · similarity 0.84)
- SSDiff: Spatial-spectral Integrated Diffusion Model for Remote Sensing Pansharpening (NeurIPS 2024 · 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.