Research map / Image generation / Super-resolution
MFTN: A Multi-scale Feature Transfer Network Based on IMatchFormer 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 MFTN, a multi-scale feature transfer network using IMatchFormer cross-modal transformers and spectral-aware aggregation to fuse LR-HSI with HR-MSI for hyperspectral super-resolution.
Problem
Hyperspectral image super-resolution (HISR) fusing a low-resolution HSI with a high-resolution MSI often yields spectral distortion and blurred spatial texture because of the feature gap between the two modalities.
Approach
Extract three-scale features from LR-HSI and degraded HR-MSI, match them with cross-modal multi-head attention (IMatchFormer) at each scale to transfer HR-MSI detail features, and aggregate them with SAAM/SAM modules that inject shallow LR-HSI spectral features.
Key results
- On Pavia Center, MFTN reaches PSNR 40.98 dB and SAM 4.20, versus 39.61 dB for the strongest baseline HyperRefiner.
- On Chikusei, MFTN achieves PSNR 42.00 dB and ERGAS 2.88, the best among the compared methods.
- Ablation on Pavia Center shows using IMatchFormer at all three scales (PSNR 40.98) outperforms single-scale variants (1x: 25.36, 2x: 34.81, 4x: 40.52).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MFTN instead of HyperRefiner | Hyperspectral image super-resolution · Pavia Center | PSNR: 39.61 -> 40.98 | Helped |
| use MFTN instead of HyperRefiner | Hyperspectral image super-resolution · Botswana | PSNR: 39.79 -> 41.97 | Helped |
| use MFTN instead of HyperPNN | Hyperspectral image super-resolution · Chikusei | PSNR: 41.57 -> 42.00 | Helped |
| add IMatchFormer at all three scales instead of using none (ablation) | MFTN · Pavia Center | PSNR: 24.82 -> 40.98 | Helped |
| add IMatchFormer at 2× and 4× scales to the 1× configuration (ablation) | MFTN · Pavia Center | PSNR: 25.36 -> 40.98 | Helped |
| use 8 attention heads instead of 1 (ablation) | IMatchFormer in MFTN · Pavia Center | PSNR: 38.49 -> 40.98 | Helped |
| increase the attention head count from 8 to 16 (ablation) | IMatchFormer in MFTN · Pavia Center | PSNR: 40.98 -> 39.95 | Hurt |
| add SAAM instead of replacing it with a sum operation and 4 residual blocks (ablation) | MFTN · Pavia Center | PSNR: 40.63 -> 40.98 | Helped |
| add SAM to SAAM (ablation) | MFTN · Pavia Center | PSNR: 40.65 -> 40.98 | Helped |
Limitations
- Runtime is not the lowest among compared methods (e.g., DARN and HyperDSNet are faster).
- Evaluated only on three public datasets with simulated degradation via Wald's protocol.
Experimental setup
Datasets: Pavia Center Botswana Chikusei
Benchmarks: Pavia Center Botswana Chikusei
Metrics: SCC SAM ERGAS UIQI SSIM PSNR
Baselines: PCA GFPCA DARN HSRNet HyperPNN SSFCNN SSRNet HyperDSNet HyperRefiner
Builds on: Transformer deformable convolution VGG-19 perceptual loss
Closest papers by meaning
- Learning Multi-Modal Cross-Scale Deformable Transformer Network for Unregistered Hyperspectral Image Super-resolution (AAAI 2024 · similarity 0.89)
- S2CycleDiff: Spatial-Spectral-Bilateral Cycle-Diffusion Framework for Hyperspectral Image Super-resolution (AAAI 2024 · similarity 0.87)
- EigenSR: Eigenimage-Bridged Pre-Trained RGB Learners for Single Hyperspectral Image Super-Resolution (AAAI 2025 · similarity 0.86)
- Enhancing Hyperspectral Images via Diffusion Model and Group-Autoencoder Super-resolution Network (AAAI 2024 · similarity 0.84)
- HIR-Diff: Unsupervised Hyperspectral Image Restoration Via Improved Diffusion Models (CVPR 2024 · similarity 0.84)
- MFTN: Multi-Level Feature Transfer Network Based on MRI-Transformer for MR Image Super-resolution (AAAI 2024 · similarity 0.84)
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.