Research map / 3D vision / Depth and stereo
Spectral Compressive Imaging via Chromaticity-Intensity Decomposition
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 a chromaticity-intensity decomposition of hyperspectral images for dual-camera CASSI and a physics-inspired unfolding network (CIDNet) with a spectral TopK Transformer and a learned spatially adaptive noise estimator.
Problem
Coded aperture snapshot spectral imaging (CASSI) reconstruction is severely ill-posed because spatial and spectral information are entangled in the measurement, and the captured radiance depends on illumination, making recovery of lighting-invariant spectral reflectance difficult.
Approach
Factorize the HSI into a spatially smooth intensity map (approximated by a PAN/RGB image in a dual-camera setup) and a spectrally sparse, illumination-invariant chromaticity cube; reconstruct chromaticity with an unfolding network using a hybrid spatial-spectral Transformer (Swin spatial encoder, TopK spectral decoder) and a dual noise estimation module that predicts spatially varying noise maps p
Key results
- CIDNet-9stg achieves 44.12 dB average PSNR and 0.991 SSIM on the 10 KAIST simulation scenes, versus 43.69 dB for In2SET-9stg.
- Chromaticity reconstruction reaches an average PSNR of 35.81 dB for CIDNet-7stg/9stg-level variants (best chromaticity metric reported), versus 28.03 dB for In2SET-9stg intensity/chromaticity baseline numbers in the table.
- Ablation: adding intensity-guided mask raises Base-1 PSNR from 35.77 to 40.83 dB; HSST and DNEM add a further 1.47 and 0.21 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the full CIDNet-9stg method | CASSI reconstruction · KAIST simulation | PSNR: 43.69 -> 44.12 | Helped |
| add intensity-guided mask (ablation) | Base-1 · KAIST simulation | PSNR: 35.77 -> 40.83 | Helped |
| add HSST (ablation) | Base-1 + intensity · KAIST simulation | PSNR: 40.83 -> 42.30 | Helped |
| add DNEM (ablation) | Base-1 + intensity + HSST · KAIST simulation | PSNR: 42.30 -> 42.51 | Helped |
| use intensity-guided mask (ablation) | ADMM · KAIST simulation | PSNR: 24.80 -> 37.09 | Helped |
| use intensity-guided mask (ablation) | MST · KAIST simulation | PSNR: 34.26 -> 37.28 | Helped |
| use intensity-guided mask (ablation) | DAUHST · KAIST simulation | PSNR: 37.21 -> 42.64 | Helped |
| add an attention module (ablation) | CIDNet-3stg Base-2 · KAIST simulation | PSNR: 40.95 -> 42.51 | Helped |
| use encoder-decoder LWSA-TKSA instead of parallel LWSA+TKSA (ablation) | CIDNet-3stg · KAIST simulation | PSNR: 42.36 -> 42.51 | Helped |
| use WSSA in encoder and decoder (ablation) | CIDNet-3stg · KAIST simulation | PSNR: 40.95 -> 41.98 | Helped |
| use TKSA in encoder and decoder (ablation) | CIDNet-3stg · KAIST simulation | PSNR: 40.95 -> 42.05 | Helped |
Limitations
- Requires a known intensity map obtained from a dual-camera setup (PAN or RGB), not applicable directly to regular single-camera CASSI.
- Chromaticity normalization assumes PAN-intensity equivalence which is approximate.
Experimental setup
Datasets: CAVE KAIST
Benchmarks: KAIST simulation dataset (10 scenes)
Metrics: PSNR SSIM
Baselines: DeSCI GAP-Net MST-L DAUHST-9stg SSR-9stg PIDS In2SET-9stg
Builds on: half-quadratic splitting Swin Transformer MST DAUHST
Closest papers by meaning
- Phy-CoSF: Physics-Guided Continuous Spectral Fields Reconstruction and Spectral Super-Resolution for Snapshot Compressive Imaging (ICML 2026 · similarity 0.89)
- SPECAT: SPatial-spEctral Cumulative-Attention Transformer for High-Resolution Hyperspectral Image Reconstruction (CVPR 2024 · similarity 0.88)
- Improving Spectral Snapshot Reconstruction with Spectral-Spatial Rectification (CVPR 2024 · similarity 0.88)
- FRN: Fractal-Based Recursive Spectral Reconstruction Network (NeurIPS 2025 · similarity 0.87)
- Dual-Window Multiscale Transformer for Hyperspectral Snapshot Compressive Imaging (AAAI 2024 · similarity 0.87)
- Detail Matters: Mamba-Inspired Joint Unfolding Network for Snapshot Spectral Compressive Imaging (AAAI 2025 · 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.