Research map / Image generation / Super-resolution
Enhancing Hyperspectral Images via Diffusion Model and Group-Autoencoder Super-resolution Network
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 DMGASR, a two-stage hyperspectral image super-resolution model that trains a diffusion model (SR3-style) in the latent space of a spectral-grouping autoencoder (GAE).
Problem
Hyperspectral image super-resolution struggles to model spectral-spatial relations, and applying diffusion models directly to high-dimensional HSI causes convergence difficulties and slow inference.
Approach
Group adjacent overlapping spectral bands and encode each group into a latent variable list with an autoencoder having an asymmetric decoder; run the conditional diffusion SR model (SR3 architecture, LR latent concatenated as condition) over this small latent list instead of full bands or band-by-band.
Key results
- On PaviaC at scale 2, the method reports MPSNR 34.491 vs. 33.326 for the second-best GELIN.
- On Chikusei at scale 4, MPSNR 32.248 vs. 31.189 for MCNet (the next best), and on Harvard at scale 4, 43.132 vs. 42.673 for GELIN.
- Inference on PaviaC at 64x64 takes 3.752 s/img versus 13.629 s/img for the band-by-band diffusion variant (Diff-PB).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add global decoding to the autoencoder (ablation) | DMGASR · PaviaC scales 2, 3, and 4 | MPSNR: 33.761 -> 34.491; 29.685 -> 30.035; 27.624 -> 27.928 | Helped |
| Add spectral grouping (ablation) | DMGASR · PaviaC scales 2, 3, and 4 | MPSNR: 32.442 -> 34.491; 29.469 -> 30.035; 27.432 -> 27.928 | Helped |
| Add the GAE instead of diffusion-only partial-band training (ablation) | DMGASR · PaviaC scales 2, 3, and 4 | MPSNR: 31.319 -> 34.491; 29.172 -> 30.035; 27.262 -> 27.928 | Helped |
| Add the GAE instead of diffusion-only full-band training (ablation) | DMGASR · PaviaC scales 2, 3, and 4 | MPSNR: 10.948 -> 34.491; 10.920 -> 30.035; 10.906 -> 27.928 | Helped |
| Use 16 subgroups and 4 overlap bands rather than other tested subgroup settings (ablation) | DMGASR · PaviaC scale 4 | MPSNR: 27.784; 27.928; 27.759; 27.736 | Helped |
| Use DMGASR rather than GELIN | HSI super-resolution · Harvard scale 2 | MPSNR: 47.024 -> 46.953 | Hurt |
| Use DMGASR rather than the strongest listed baseline | HSI super-resolution · PaviaC scales 2, 3, and 4; Chikusei scales 2, 3, and 4; Harvard scales 3 and 4 | MPSNR: 33.326 -> 34.491; 29.611 -> 30.035; 27.854 -> 27.928; 37.747 -> 38.748; 33.796 -> 34.192; 31.189 -> 32.248; 43.653 -> 44 | Helped |
Surprising findings
- Direct full-band diffusion without the autoencoder collapses (MPSNR about 10.9 dB across scales), showing the latent compression is essential for convergence.
- The method has the worst SAM on several settings (e.g., 7.406 at PaviaC scale 4 vs. 6.265 for GELIN), which the authors attribute to segmented band processing.
Limitations
- SAM spectral fidelity is not consistently best and is worse than several baselines at some scales.
- Compute hardware and training duration are not reported.
- Evaluated only on three HSI datasets with a single diffusion backbone (SR3).
Experimental setup
Datasets: PaviaC Chikusei Harvard
Benchmarks: PaviaC Chikusei Harvard
Metrics: MPSNR MSSIM CC RMSE SAM ERGAS
Baselines: Bicubic EDSR GDRRN SSPSR MCNet CEGATSR GELIN
Builds on: SR3 latent diffusion models GELIN SSPSR
Closest papers by meaning
- HIR-Diff: Unsupervised Hyperspectral Image Restoration Via Improved Diffusion Models (CVPR 2024 · similarity 0.88)
- S2CycleDiff: Spatial-Spectral-Bilateral Cycle-Diffusion Framework for Hyperspectral Image Super-resolution (AAAI 2024 · similarity 0.88)
- EigenSR: Eigenimage-Bridged Pre-Trained RGB Learners for Single Hyperspectral Image Super-Resolution (AAAI 2025 · similarity 0.87)
- Hyperspectral Pansharpening via Diffusion Models with Iteratively Zero-Shot Guidance (CVPR 2025 · similarity 0.86)
- Unmixing Diffusion for Self-Supervised Hyperspectral Image Denoising (CVPR 2024 · similarity 0.86)
- Unmixing Before Fusion: A Generalized Paradigm for Multi-Source-based Hyperspectral Image Synthesis (CVPR 2024 · 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.