Research map / Image generation / Image restoration
FRN: Fractal-Based Recursive Spectral Reconstruction 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 FRN, a fractal-inspired recursive spectral reconstruction network that generates HSI bands progressively from RGB using a band-aware Mamba atomic generator, achieving top PSNR on CAVE and Harvard with 0.30M parameters.
Problem
RGB-to-hyperspectral spectral reconstruction is ill-posed, and existing one-shot transformer/CNN methods integrate full RGB spectral information in a brute-force way with high parameter and compute cost.
Approach
Treats spectral reconstruction as a recursive, fractal-like coarse-to-fine process in which a shared atomic generator (BAMamba, a VMamba-style SSM with a band-aware mask from the discretized Delta-A coefficient) predicts two new bands per level from neighboring bands and RGB cues.
Key results
- On CAVE, FRN reaches 41.05 dB PSNR and 0.9900 SSIM, outperforming the second-best (MSFN, 39.84 dB) by about 1.2 dB, with 0.30M parameters.
- On Harvard, FRN reaches 42.88 dB PSNR, 0.23 dB above the second-best method (MSFN, 42.65 dB).
- Ablation: removing the band-aware mask (vanilla VMamba) drops CAVE PSNR from 41.05 to 39.75 dB; a one-step (no recursion) variant gives 39.86 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace MSFN with FRN | Spectral reconstruction · CAVE test | PSNR: 39.8430 -> 41.0522 | Helped |
| replace MSFN with FRN | Spectral reconstruction · CAVE test | RMSE: 4.0372 -> 3.6243 | Helped |
| replace MSFN with FRN | Spectral reconstruction · Harvard test | PSNR: 42.6455 -> 42.8762 | Helped |
| replace MSFN with FRN | Spectral reconstruction · Harvard test | RMSE: 2.9916 -> 2.8933 | Helped |
| add band-aware mask (vanilla VMamba to α=0.5) (ablation) | FRN · CAVE | PSNR: 39.7482 -> 41.0522 | Helped |
| set suppression threshold to α=0.2 or α=0.3 instead of α=0.5 (ablation) | FRN · CAVE | PSNR: 41.0522 -> 39.9022; 40.3285 | Hurt |
| set suppression threshold to α=0.7 or α=0.8 instead of α=0.5 (ablation) | FRN · CAVE | PSNR: 41.0522 -> 40.1822; 37.4822 | Hurt |
| replace recursive reconstruction with one-step RGB-to-HSI reconstruction (ablation) | FRN · CAVE | PSNR: 41.0522 -> 39.8644 | Hurt |
| use M=2 or M=3 rather than M=5 recursive levels (ablation) | FRN · CAVE | PSNR: 40.2210; 40.6382 -> 41.0522 | Hurt |
| exclude RGB image from each reconstruction level (ablation) | FRN · CAVE | PSNR: 38.6244 -> 41.0522 | Hurt |
| use S=2 or S=3 reference spectral channels rather than S=4 (ablation) | FRN · CAVE | PSNR: 40.3981; 40.8286 -> 41.0522 | Hurt |
| use S=5 reference spectral channels rather than S=4 (ablation) | FRN · CAVE | PSNR: 41.0522 -> 40.8744 | Hurt |
Surprising findings
- Removing RGB inputs from each recursive level drops CAVE PSNR to 38.62 dB, showing RGB cues are essential at every level.
- An overly high suppression threshold (alpha=0.8) sharply degrades performance (37.48 dB), indicating informative features get suppressed.
Limitations
- Evaluated only on two datasets (CAVE and Harvard) with 31-band HSIs.
- Band-aware threshold alpha requires empirical tuning; performance is sensitive to it.
- Compares against a limited set of seven baselines and does not report FLOPs or runtime in the text.
Experimental setup
Datasets: CAVE Harvard
Benchmarks: CAVE Harvard
Metrics: PSNR RMSE UIQI SSIM
Baselines: HSACS SSRNet HSRNet AWAN MST++ LTRN MSFN
Builds on: VMamba MST++
Closest papers by meaning
- Unraveling Metameric Dilemma for Spectral Reconstruction: A High-Fidelity Approach via Semi-Supervised Learning (NeurIPS 2025 · similarity 0.88)
- Spectral Compressive Imaging via Chromaticity-Intensity Decomposition (NeurIPS 2025 · similarity 0.87)
- Hyperspectral Image Reconstruction via Combinatorial Embedding of Cross-Channel Spatio-Spectral Clues (AAAI 2024 · similarity 0.86)
- SPECAT: SPatial-spEctral Cumulative-Attention Transformer for High-Resolution Hyperspectral Image Reconstruction (CVPR 2024 · similarity 0.86)
- Improving Spectral Snapshot Reconstruction with Spectral-Spatial Rectification (CVPR 2024 · similarity 0.86)
- EigenSR: Eigenimage-Bridged Pre-Trained RGB Learners for Single Hyperspectral Image Super-Resolution (AAAI 2025 · similarity 0.85)
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.