Research map / Image generation / Image restoration
Automatic Spectral Calibration of Hyperspectral Images: Method, Dataset and Benchmark
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.
Creates the BJTU-UVA hyperspectral spectral-calibration dataset (765 pairs, expanded to 7650 with 10 measured illuminations) and proposes SIT, a U-shaped spectral transformer with an illumination attention module inspired by Gray-World, for learning-based HSI calibration without a physical reference.
Problem
Calibrating hyperspectral images to remove illumination effects normally requires a physical white reference, which causes occlusion, manual operations, and limits camera mobility; no learning-based method existed for this task.
Approach
Learn to infer scene reflectance directly from an uncalibrated HSI using a U-shaped spectral transformer whose attention is augmented by an illumination attention branch that mimics the Gray-World global-mean assumption.
Key results
- SIT reaches 26.3 PSNR and 3.1 SAM on BJTU-UVA test (full spectrum), versus 25.7 PSNR / 3.6 SAM for HCANet and 23.3 PSNR for Gray-World.
- On BJTU-UVA-E full spectrum test, SIT achieves 39.1 PSNR and 2.0 SAM, the best among compared methods.
- Ablation on BJTU-UVA shows the illumination attention alone (IA) lifts test PSNR from 22.1 (no SA/IA) to 25.8, and combining IA and SA gives 26.3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SIT instead of HCANet | Automatic spectral calibration · BJTU-UVA full-spectrum test | PSNR: 25.7 -> 26.3 | Helped |
| use SIT instead of Gray-World | Automatic spectral calibration · BJTU-UVA full-spectrum test | SAM: 3.1 -> 3.1 | No clear effect |
| use SIT instead of HCANet | Automatic spectral calibration · BJTU-UVA-E full-spectrum test | PSNR: 37.3 -> 39.1 | Helped |
| use SIT instead of HCANet | Automatic spectral calibration · BJTU-UVA-E full-spectrum test | SAM: 2.5 -> 2.0 | Helped |
| use SIT instead of HCANet | Automatic spectral calibration · BJTU-UVA 31-channel test | PSNR: 26.3 -> 26.7 | Helped |
| use SIT instead of HCANet | Automatic spectral calibration · BJTU-UVA 31-channel test | SAM: 3.4 -> 2.7 | Helped |
| use SIT instead of HCANet | Automatic spectral calibration · BJTU-UVA-E 31-channel test | PSNR: 38.1 -> 41.1 | Helped |
| use SIT instead of HCANet | Automatic spectral calibration · BJTU-UVA-E 31-channel test | SAM: 2.5 -> 1.9 | Helped |
| add illumination attention to spectral attention (ablation) | SIT · BJTU-UVA full-spectrum test | PSNR: 25.7 -> 26.3 | Helped |
| add illumination attention to spectral attention (ablation) | SIT · BJTU-UVA full-spectrum test | SAM: 3.6 -> 3.1 | Helped |
| add spectral attention to illumination attention (ablation) | SIT · BJTU-UVA full-spectrum test | PSNR: 25.8 -> 26.3 | Helped |
| add spectral attention to illumination attention (ablation) | SIT · BJTU-UVA full-spectrum test | SAM: 4.3 -> 3.1 | Helped |
Surprising findings
- Gray-World is competitive with or better than learned methods on the original all-natural-illumination BJTU-UVA dataset in SAM scores.
- Learning-based methods perform much better on the expanded dataset than on the original, while Gray-World does not improve.
Limitations
- Low-light scenes and areas show larger calibration errors, especially in infrared bands.
- Gray-World underperforms learned methods on the synthetic-illumination expanded dataset.
- Assumes uniform global illumination.
Experimental setup
Datasets: BJTU-UVA BJTU-UVA-E
Benchmarks: BJTU-UVA BJTU-UVA-E
Metrics: PSNR SAM RMSE ERGAS
Baselines: Gray-World DivIll SERT HCANet
Builds on: HCANet Gray-World U-Net DivIll
Closest papers by meaning
- EigenSR: Eigenimage-Bridged Pre-Trained RGB Learners for Single Hyperspectral Image Super-Resolution (AAAI 2025 · similarity 0.87)
- Improving Spectral Snapshot Reconstruction with Spectral-Spatial Rectification (CVPR 2024 · similarity 0.86)
- S2CycleDiff: Spatial-Spectral-Bilateral Cycle-Diffusion Framework for Hyperspectral Image Super-resolution (AAAI 2024 · similarity 0.84)
- Unmixing Before Fusion: A Generalized Paradigm for Multi-Source-based Hyperspectral Image Synthesis (CVPR 2024 · similarity 0.84)
- SPECAT: SPatial-spEctral Cumulative-Attention Transformer for High-Resolution Hyperspectral Image Reconstruction (CVPR 2024 · similarity 0.84)
- Illumination Spectrum Estimation for Multispectral Images via Surface Reflectance Modeling and Spatial-Spectral Feature Generation (CVPR 2025 · 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.