Research map / Image generation / Image restoration
HIR-Diff: Unsupervised Hyperspectral Image Restoration Via Improved Diffusion Models
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 HIR-Diff, an unsupervised hyperspectral image restoration method that factorizes the HSI into a reduced image (restored by a pre-trained diffusion model with a TV-guided loss and exponential noise schedule) and an SVD/RRQR-estimated coefficient matrix.
Problem
Hyperspectral image restoration (denoising, noisy super-resolution, noisy inpainting) suffers from handcrafted priors in model-based methods and poor generalization and data scarcity in supervised deep learning methods, while existing self-supervised diffusion methods are slow.
Approach
Factor the clean HSI as a product of a low-spectral-dimension reduced image and a coefficient matrix; estimate the coefficient matrix via SVD with RRQR-based band selection, and sample the reduced image from a pre-trained RGB diffusion model guided by a data-fidelity plus TV loss, using a new exponential noise schedule to speed up DDIM sampling.
Key results
- HIR-Diff achieves 42.85 PSNR on WDC Mall denoising at sigma=30, best among compared methods, in 17 s versus 3132 s for DDS2M.
- Exponential noise schedule yields 36.01 PSNR denoising at 20 sampling steps vs 34.34 for linear schedule, enabling about 5x acceleration.
- On noisy inpainting (mask rate 0.9, WDC Mall), HIR-Diff reaches 35.08 PSNR vs 32.35 for DDS2M.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HIR-Diff rather than the strongest listed baseline | HIR-Diff · HSI denoising, σ=30; WDC mall, Houston, Salinas | PSNR: 43.77 -> 42.85; 38.50 -> 38.13; 44.76 -> 43.79 | Mixed |
| Use HIR-Diff rather than the strongest listed baseline | HIR-Diff · Noisy HSI super-resolution ×2; WDC mall, Houston, Salinas | PSNR: 34.55 -> 36.67; 30.52 -> 31.83; 36.29 -> 39.44 | Helped |
| Use HIR-Diff rather than the strongest listed baseline | HIR-Diff · Noisy HSI inpainting, mask rate 0.7; WDC mall, Houston, Salinas | PSNR: 38.32 -> 37.90; 31.91 -> 32.42; 40.49 -> 40.70 | Mixed |
| Add the data fidelity term and TV regularization to the guidance function (ablation) | HIR-Diff · Denoising, super-resolution, inpainting | PSNR: 10.12 -> 36.14; 11.49 -> 30.70; 11.13 -> 31.71 | Helped |
| Add TV regularization to data-fidelity guidance (ablation) | HIR-Diff · Denoising, super-resolution, inpainting | PSNR: 34.99 -> 36.14; 27.10 -> 30.70; 29.90 -> 31.71 | Helped |
| Use the proposed reduced-image decomposition rather than SVD-only or least-squares alternatives (ablation) | HIR-Diff · WDC mall denoising, super-resolution, inpainting | PSNR: 28.88 -> 40.73; 27.86 -> 34.79; 28.16 -> 37.02 | Helped |
| Select reduced-image bands using RRQR rather than equally spaced indices (ablation) | HIR-Diff · Salinas denoising, super-resolution, inpainting | PSNR: 27.85 -> 44.27; 11.02 -> 39.68; 34.54 -> 41.02 | Helped |
| Replace the linear noise schedule with the exponential schedule (ablation) | HIR-Diff · Denoising, super-resolution, inpainting, 20 sampling steps | PSNR: 34.34 -> 36.01; 29.47 -> 30.68; 29.96 -> 31.19 | Helped |
| Replace the cosine noise schedule with the exponential schedule (ablation) | HIR-Diff · Denoising, super-resolution, inpainting, 20 sampling steps | PSNR: 34.61 -> 36.01; 30.25 -> 30.68; 30.17 -> 31.19 | Helped |
| Increase exponential-schedule sampling from 20 to 100 steps (ablation) | HIR-Diff · Denoising, super-resolution, inpainting | PSNR: 36.01 -> 36.14; 30.68 -> 30.70; 31.19 -> 31.71 | Helped |
Surprising findings
- Linear and cosine noise schedules yield fluctuating PSNR that fails to converge even with more steps or stronger guidance.
- Denoising at sigma=30 on WDC Mall: DDS2M reaches higher PSNR than HIR-Diff at mask rates 0.7 (38.32 vs 37.90) in inpainting, showing the method is not uniformly best.
Limitations
- Relies on a diffusion model pre-trained on RGB remote sensing images, so the reduced image distribution depends on that prior.
- Hyperparameters (lambda, beta, k, epsilon) must be set; guidance and schedule rely on tuning.
- Inpainting at high mask rate is sometimes outperformed by DDS2M on some datasets.
Experimental setup
Datasets: Washington DC Mall Houston Salinas ICVL CAVE
Benchmarks: Washington DC Mall Houston Salinas
Metrics: PSNR SSIM running time (s)
Baselines: BM4D NGMeet ETPTV T3SC SST SERT DDS2M DIP2d DIP3d TRPCA TRLRF S2NTNN
Builds on: DDRM DDS2M DDIM NGMeet pansharpening diffusion method (Wang et al.)
Closest papers by meaning
- Unmixing Diffusion for Self-Supervised Hyperspectral Image Denoising (CVPR 2024 · similarity 0.90)
- EMR-Diff: Edge-aware Multimodal Residual Diffusion Model for Hyperspectral Image Super-resolution (CVPR 2026 · similarity 0.90)
- Real Noise Decoupling for Hyperspectral Image Denoising (AAAI 2026 · similarity 0.89)
- Enhancing Hyperspectral Images via Diffusion Model and Group-Autoencoder Super-resolution Network (AAAI 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 0.86)
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.