Research map / Image generation / Super-resolution
S2CycleDiff: Spatial-Spectral-Bilateral Cycle-Diffusion Framework for Hyperspectral Image Super-resolution
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 S2CycleDiff, a conditional cycle-diffusion model that performs spatial and spectral super-resolution bilaterally with guided pyramid denoising and a fusion block to fuse HSI and MSI into high-resolution hyperspectral images.
Problem
Hyperspectral image super-resolution (fusing low-resolution HSI with high-resolution MSI) where existing methods learn a joint spatial-spectral mapping and fail to preserve the spectral distribution of LrHSI and the spatial distribution of HrMSI.
Approach
A bilateral conditional cycle-diffusion backbone runs a spatial branch (conditioned on LrHSI, guided by HrMSI) and a spectral branch (conditioned on HrMSI, guided by LrHSI) in parallel, with cross-attention detail-injection blocks in pyramid denoisers, consistency losses via SRF/PSF, and a complementary fusion block.
Key results
- On CAVE, S2CycleDiff reaches PSNR 43.93 and SAM 2.79 at x4, beating PSRT (43.48 / 2.88).
- On Chikusei, PSNR 43.88 vs. 42.9 for the best deep baseline (SSR-NET 39.53 in the table), with SAM 1.28 and ERGAS 1.34.
- At x16 on CAVE, PSNR 39.92 vs. 39.64 for LAGC-NET, the next best.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use bilateral spatial and spectral super-resolution branches instead of only the spectral branch (ablation) | S2CycleDiff · CAVE | PSNR: 41.0376 -> 43.9264 | Helped |
| Use bilateral spatial and spectral super-resolution branches instead of only the spatial branch (ablation) | S2CycleDiff · CAVE | PSNR: 41.3527 -> 43.9264 | Helped |
| Use bilateral spatial and spectral super-resolution branches instead of only the spectral branch (ablation) | S2CycleDiff · Chikusei | PSNR: 41.6815 -> 43.8817 | Helped |
| Use bilateral spatial and spectral super-resolution branches instead of only the spatial branch (ablation) | S2CycleDiff · Chikusei | PSNR: 41.4683 -> 43.8817 | Helped |
| Use bilateral spatial and spectral super-resolution branches instead of only the spectral branch (ablation) | S2CycleDiff · Pavia Center | PSNR: 40.5205 -> 43.3698 | Helped |
| Use bilateral spatial and spectral super-resolution branches instead of only the spatial branch (ablation) | S2CycleDiff · Pavia Center | PSNR: 41.9572 -> 43.3698 | Helped |
| Add guiding factors to the detail injection module instead of removing them (ablation) | S2CycleDiff · CAVE | PSNR: 36.2690 -> 43.9264 | Helped |
| Use cross-attention to integrate guiding factors instead of direct concatenation (ablation) | S2CycleDiff · CAVE | PSNR: 43.4294 -> 43.9264 | Helped |
| Add guiding factors to the detail injection module instead of removing them (ablation) | S2CycleDiff · Chikusei | PSNR: 29.1468 -> 43.8817 | Helped |
| Use cross-attention to integrate guiding factors instead of direct concatenation (ablation) | S2CycleDiff · Chikusei | PSNR: 42.9069 -> 43.8817 | Helped |
| Add guiding factors to the detail injection module instead of removing them (ablation) | S2CycleDiff · Pavia Center | PSNR: 39.3987 -> 43.3698 | Helped |
| Use cross-attention to integrate guiding factors instead of direct concatenation (ablation) | S2CycleDiff · Pavia Center | PSNR: 42.4301 -> 43.3698 | Helped |
Limitations
- Only three datasets and a fixed degradation pipeline (Wald's protocol, factor 4 main setting).
- Diffusion sampling with T=2000 steps implies high inference cost; no runtime analysis reported.
Experimental setup
Datasets: CAVE Chikusei Pavia Center
Benchmarks: CAVE Chikusei Pavia Center
Metrics: PSNR SAM ERGAS SSIM
Baselines: GSA FUSE CNMF SSR-NET MoG-DCN LAGC-NET LightNet PSRT
Builds on: DDPM PSRT MoG-DCN
Closest papers by meaning
- EigenSR: Eigenimage-Bridged Pre-Trained RGB Learners for Single Hyperspectral Image Super-Resolution (AAAI 2025 · similarity 0.88)
- HIR-Diff: Unsupervised Hyperspectral Image Restoration Via Improved Diffusion Models (CVPR 2024 · similarity 0.88)
- Enhancing Hyperspectral Images via Diffusion Model and Group-Autoencoder Super-resolution Network (AAAI 2024 · similarity 0.88)
- MFTN: A Multi-scale Feature Transfer Network Based on IMatchFormer for Hyperspectral Image Super-Resolution (ICML 2024 · similarity 0.87)
- Improving Spectral Snapshot Reconstruction with Spectral-Spatial Rectification (CVPR 2024 · similarity 0.87)
- Learning Multi-Modal Cross-Scale Deformable Transformer Network for Unregistered Hyperspectral Image Super-resolution (AAAI 2024 · similarity 0.87)
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.