Research map / Image generation / Image restoration
Self-Learning Hyperspectral and Multispectral Image Fusion via Adaptive Residual Guided Subspace Diffusion Model
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 ARGS-Diff, a self-learning subspace diffusion model for hyperspectral-multispectral fusion that trains lightweight spectral and spatial diffusion networks solely on the observed LR-HSI and HR-MSI, with an adaptive residual guided module refining components at each sampling step.
Problem
Hyperspectral and multispectral image fusion (HSI-MSI) to produce high-resolution hyperspectral images when paired training data is scarce, while diffusion-based fusion methods are slow and memory-heavy.
Approach
Decompose the HR-HSI into a spectral basis and a reduced coefficient, learn each with a lightweight diffusion network (a fully connected spectral net and a UNet-like spatial net) trained only on the observed images, and stabilize joint posterior sampling with a residual-guided refinement (ARGM) at each step.
Key results
- On Pavia, ARGS-Diff reaches 42.33 dB PSNR, 1.31 dB above the second-best MIAE (41.02 dB), in 12 s.
- On Chikusei and KSC, ARGS-Diff reports 41.90 dB and 43.63 dB PSNR, outperforming the best baselines by 1.37 dB and 1.27 dB respectively.
- ARGS-Diff uses 21.85M parameters and 2.11 GB memory versus 391M/6.21 GB for PLRDiff and 606M/23.34 GB for S2CycleDiff on Pavia.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add ARGM (ablation) | ARGS-Diff · Pavia | PSNR: 41.86 -> 42.33 | Helped |
| add ARGM (ablation) | ARGS-Diff · Pavia | SAM: 2.75 -> 2.64 | Helped |
| add ARGM (ablation) | ARGS-Diff · Chikusei | PSNR: 41.33 -> 41.90 | Helped |
| add ARGM (ablation) | ARGS-Diff · Chikusei | SAM: 1.89 -> 1.77 | Helped |
| add ARGM (ablation) | ARGS-Diff · KSC | PSNR: 43.02 -> 43.63 | Helped |
| add ARGM (ablation) | ARGS-Diff · KSC | SAM: 2.70 -> 2.54 | Helped |
Limitations
- Relies on a fixed subspace dimension d=8 and hyperparameters tuned on Pavia, with sensitivity shown only on that dataset
- Performance gains reported on simulated degradations with bicubic downsampling; real-data evaluation is only qualitative on Houston
Experimental setup
Datasets: Pavia University Chikusei KSC DFC2018 Houston
Benchmarks: Pavia University Chikusei KSC DFC2018 Houston
Metrics: PSNR SAM ERGAS SSIM runtime
Baselines: CNMF HySure CuCaNet MIAE PLRDiff S2CycleDiff
Builds on: DDPM diffusion posterior sampling PLRDiff S2CycleDiff Adam
Closest papers by meaning
- A Selective Re-learning Mechanism for Hyperspectral Fusion Imaging (CVPR 2025 · similarity 0.85)
- SSDiff: Spatial-spectral Integrated Diffusion Model for Remote Sensing Pansharpening (NeurIPS 2024 · similarity 0.81)
- Enhancing Hyperspectral Images via Diffusion Model and Group-Autoencoder Super-resolution Network (AAAI 2024 · similarity 0.81)
- Unmixing Diffusion for Self-Supervised Hyperspectral Image Denoising (CVPR 2024 · similarity 0.81)
- HIR-Diff: Unsupervised Hyperspectral Image Restoration Via Improved Diffusion Models (CVPR 2024 · similarity 0.80)
- Hyperspectral Pansharpening via Diffusion Models with Iteratively Zero-Shot Guidance (CVPR 2025 · similarity 0.80)
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.