Research map / Image generation / Image restoration
Unmixing Diffusion for Self-Supervised Hyperspectral Image Denoising
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.
Diff-Unmix denoises hyperspectral images self-supervised by combining a transformer-based spectral unmixing network with a pretrained RGB diffusion model that refines abundance maps via Markov chain state matching and a learned conditioning function.
Problem
Denoising hyperspectral images without paired supervised training data, where noise corrupts both spatial detail and spectral consistency.
Approach
Decompose the noisy HSI into abundance maps and endmembers via a Spectral Transformer Unmixing network, then refine the abundance maps with a pretrained unconditional diffusion model, conditioned by a self-supervised J-invariance-based function and started from a matched intermediate diffusion state; the result is recomposed with endmembers.
Key results
- On KAIST with N(0,0.2) noise, Diff-Unmix reaches 33.059 PSNR and 0.964 SSIM, versus 33.011 PSNR for HLRTF and 32.804 for DDS2M.
- On CAVE-Toy, Diff-Unmix achieves 28.046 PSNR and 0.945 SSIM, with FSIM 0.993, but DDS2M has higher PSNR (29.344).
- State matching reduces inference time from 311s to 41s in the Fig. 6 comparison (matched state t=61 vs T=500).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Diff-Unmix instead of HLRTF | HSI denoising · KAIST, N(0, 0.2) | PSNR: 33.011 -> 33.059 | Helped |
| Use Diff-Unmix instead of E3DTV | HSI denoising · KAIST, N(0, 0.2) | SSIM: 0.868 -> 0.964 | Helped |
| Use Diff-Unmix instead of HLRTF | HSI denoising · KAIST, N(0, 0.3) | PSNR: 30.340 -> 31.408 | Helped |
| Use Diff-Unmix instead of E3DTV | HSI denoising · KAIST, N(0, 0.3) | SSIM: 0.819 -> 0.902 | Helped |
| Use Diff-Unmix instead of DDS2M | HSI denoising · CAVE-Toy, N(0, 0.3) | PSNR: 29.344 -> 28.046 | Hurt |
| Use Diff-Unmix instead of DDS2M | HSI denoising · CAVE-Toy, N(0, 0.3) | SSIM: 0.844 -> 0.945 | Helped |
| Use Diff-Unmix instead of E3DTV | HSI denoising · CAVE, N(0, 0.3) | PSNR: 32.344 -> 32.714 | Helped |
| Use Diff-Unmix instead of E3DTV | HSI denoising · CAVE, N(0, 0.3) | SSIM: 0.951 -> 0.940 | Hurt |
| Add matched Markov-chain start-state selection (ablation) | Diff-Unmix · KAIST | inference time: 311s -> 41s | Helped |
| Add conditioning function Φ (ablation) | Diff-Unmix · HSI denoising, visual comparison | restoration detail quality | Helped |
Surprising findings
- On CAVE-Toy the proposed method has lower PSNR than DDS2M (28.046 vs 29.344) and a lower-than-expected SAM than several baselines, so it does not win on all metrics.
Limitations
- Can generate spurious details (over-enhancement) due to generative nature, as shown in Fig. 11.
- Inference time remains high relative to some optimization baselines (around 40s per image).
Experimental setup
Datasets: CAVE
Benchmarks: KAIST CAVE CAVE-Toy Urban
Metrics: PSNR SSIM FSIM SAM running time
Baselines: NonLRMA LRTDTV LLRSSTV TLRLSSTV LLxRGTV 3DTNN 3DTNN-FW LRTDCTV E3DTV FGSLR DIP LLRPnP
Builds on: DDPM DDRM DDS2M J-Invariance Spectral-wise attention (SST)
Closest papers by meaning
- HIR-Diff: Unsupervised Hyperspectral Image Restoration Via Improved Diffusion Models (CVPR 2024 · similarity 0.90)
- Unmixing Before Fusion: A Generalized Paradigm for Multi-Source-based Hyperspectral Image Synthesis (CVPR 2024 · similarity 0.90)
- Enhancing Unregistered Hyperspectral Image Super-Resolution via Unmixing-based Abundance Fusion Learning (CVPR 2026 · similarity 0.89)
- DiffSCI: Zero-Shot Snapshot Compressive Imaging via Iterative Spectral Diffusion Model (CVPR 2024 · similarity 0.87)
- S2CycleDiff: Spatial-Spectral-Bilateral Cycle-Diffusion Framework for Hyperspectral Image Super-resolution (AAAI 2024 · similarity 0.86)
- Enhancing Hyperspectral Images via Diffusion Model and Group-Autoencoder Super-resolution Network (AAAI 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.