Research map / Architectures and efficiency / Attention and transformers
Improving Spectral Snapshot Reconstruction with Spectral-Spatial Rectification
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 Spectral-Spatial Rectification (SSR), with window-based spectra-wise self-attention (WSSA) and a spatial rectification block (ARB), for snapshot spectral compressive imaging reconstruction.
Problem
Spectra-wise transformers for snapshot spectral compressive imaging (CASSI) reconstruction use multi-head division along the spectral dimension, which limits global spectral modeling and causes a "mean effect"; also, unique spatial degradation from masking, shifting, and compression is not addressed
Approach
Divide features into spatial windows before computing spectra-wise self-attention (WSSA) to capture global spectral similarity while keeping local differences and spatial correlation; add a large-kernel convolution modulation block for cross-window interaction and a spatial alignment block that uses high-quality bands to rectify low-quality bands.
Key results
- SSR-L reaches 40.69 dB average PSNR on 10 KAIST scenes, versus 39.57 dB for RDLUF (prior SOTA).
- SSR-S (1.73M params, 26.37 GFLOPs) reaches 39.19 dB, surpassing RDLUF at lower compute.
- Ablation: baseline 36.60 dB; adding WSSA gives 37.76 dB and adding ARB gives 39.19 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SSR-L instead of RDLUF | Spectral snapshot reconstruction · KAIST (10 scenes), Avg | PSNR: 1.12dB | Helped |
| Add WSSA to the baseline (ablation) | SSR-S · KAIST and CAVE ablation | PSNR: 1.16dB | Helped |
| Add ARB after WSSA (ablation) | SSR-S · KAIST and CAVE ablation | PSNR: 1.43dB | Helped |
| Use WSSA instead of S-MSA (ablation) | SSR-S · Self-attention scheme ablation | PSNR: 0.45dB | Helped |
| Use WSSA instead of Swin-MSA (ablation) | SSR-S · Self-attention scheme ablation | PSNR: 0.29dB | Helped |
| Use WSSA instead of HS-MSA (ablation) | SSR-S · Self-attention scheme ablation | PSNR: 0.21dB | Helped |
| Use WSSA instead of WSSA* (ablation) | SSR-S · Self-attention scheme ablation | PSNR: 0.11 | Helped |
| Share parameters across stages (ablation) | SSR-L · KAIST (10 scenes), Avg | PSNR: performance degradation | Hurt |
Surprising findings
- Multi-head attention in the spectral dimension is argued to work because of a mean effect, and the WSSA window size (8x8) gives best results, with large windows degrading performance.
- Parameter-shared SSR-L* degrades performance relative to SSR-L but still beats other methods with the fewest parameters.
Limitations
- Real-scene evaluation lacks ground truth, so quantitative comparison is not possible there.
- Performance of the parameter-shared variant degrades relative to the full model.
Experimental setup
Datasets: CAVE KAIST
Benchmarks: CAVE KAIST (10 scenes) real CASSI measurements (5 scenes)
Metrics: PSNR SSIM
Baselines: DIP-HSI TSA-Net DGSMP GAP-Net HerosNet HDNet MST CST BIRNAT DAUHST RDLUF
Builds on: MST CMB (Conv-modulated block) GAP-Net DAUHST
Closest papers by meaning
- Dual-Window Multiscale Transformer for Hyperspectral Snapshot Compressive Imaging (AAAI 2024 · similarity 0.90)
- SPECAT: SPatial-spEctral Cumulative-Attention Transformer for High-Resolution Hyperspectral Image Reconstruction (CVPR 2024 · similarity 0.89)
- Spectral Compressive Imaging via Chromaticity-Intensity Decomposition (NeurIPS 2025 · similarity 0.88)
- S2CycleDiff: Spatial-Spectral-Bilateral Cycle-Diffusion Framework for Hyperspectral Image Super-resolution (AAAI 2024 · similarity 0.87)
- Sp3ctralMamba: Physics-Driven Joint State Space Model for Hyperspectral Image Reconstruction (AAAI 2025 · similarity 0.86)
- Automatic Spectral Calibration of Hyperspectral Images: Method, Dataset and Benchmark (CVPR 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.