Research map / Architectures and efficiency / Attention and transformers
Dual-Window Multiscale Transformer for Hyperspectral Snapshot Compressive Imaging
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 DWMT, a Transformer U-Net with a dual-branch encoder and dual-window multiscale self-attention (DWM-MSA) plus a con-abs position embedding (CAPE) for hyperspectral snapshot compressive imaging reconstruction from CASSI measurements.
Problem
Reconstructing hyperspectral images from CASSI 2D snapshot measurements requires capturing long-range dependencies and local detail while keeping computational cost reasonable for Transformers.
Approach
Splits attention channels into four branches using two window sizes (8x8 and 16x16) with shuffle-based long-range attention inside windows, combined with a dual-branch U-Net encoder for pixel refinement and a hybrid conditional plus absolute position embedding.
Key results
- DWMT achieves average 36.82 dB PSNR and 0.964 SSIM on ten KAIST scenes, versus 36.12 dB for CST and 35.18 dB for MST.
- Removing DWM-MSA drops PSNR by 4.07 dB in the ablation study (32.23 vs 36.82 dB baseline without DWM-MSA setting).
- CAPE improves PSNR by 0.31 dB over APE and 0.51 dB over CPE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DWMT instead of CST | Hyperspectral reconstruction · KAIST (10-scene average) | PSNR (dB): 36.12 -> 36.82 | Helped |
| use DWMT instead of CST | Hyperspectral reconstruction · KAIST (10-scene average) | SSIM: 0.957 -> 0.964 | Helped |
| add the fine-pixel-refinement enhancement branch (ablation) | DWMT · KAIST simulation | PSNR (dB): 32.23 -> 32.34 | Helped |
| add DWM-MSA instead of using no self-attention (ablation) | DWMT with the enhancement branch · KAIST simulation | PSNR (dB): 32.34 -> 36.82 | Helped |
| use DWM-MSA rather than SAH-MSA (ablation) | DWMT · KAIST simulation | PSNR (dB): 36.46 -> 36.82 | Helped |
| add CAFM to the encoder with DWM-MSA (ablation) | DWMT · KAIST simulation | PSNR (dB): 36.55 -> 36.82 | Helped |
| use dual-window attention instead of a single 8×8 window (ablation) | DWMT · KAIST simulation | PSNR (dB): 36.66 -> 36.82 | Helped |
| use dual-window attention instead of a single 16×16 window (ablation) | DWMT · KAIST simulation | PSNR (dB): 36.40 -> 36.82 | Helped |
| use CAPE instead of APE (ablation) | DWMT · KAIST simulation | PSNR (dB): 36.51 -> 36.82 | Helped |
| use CAPE instead of CPE (ablation) | DWMT · KAIST simulation | PSNR (dB): 36.31 -> 36.82 | Helped |
Limitations
- Higher parameter count (14.48M) than MST (2.03M) and CST (3.00M)
- Evaluation limited to CAVE/KAIST simulation and one real CASSI dataset
Experimental setup
Datasets: CAVE KAIST
Benchmarks: KAIST (ten scenes) real CASSI HSI data
Metrics: PSNR SSIM Params FLOPs
Baselines: GAP-TV DeSCI TSA-Net HDNet MST CST
Builds on: Swin Transformer Half-Shuffle Transformer CST MST conditional position embedding
Closest papers by meaning
- Improving Spectral Snapshot Reconstruction with Spectral-Spatial Rectification (CVPR 2024 · similarity 0.90)
- SPECAT: SPatial-spEctral Cumulative-Attention Transformer for High-Resolution Hyperspectral Image Reconstruction (CVPR 2024 · similarity 0.87)
- In2SET: Intra-Inter Similarity Exploiting Transformer for Dual-Camera Compressive Hyperspectral Imaging (CVPR 2024 · similarity 0.87)
- Learning Multi-Modal Cross-Scale Deformable Transformer Network for Unregistered Hyperspectral Image Super-resolution (AAAI 2024 · similarity 0.87)
- Spectral Compressive Imaging via Chromaticity-Intensity Decomposition (NeurIPS 2025 · similarity 0.87)
- Sp3ctralMamba: Physics-Driven Joint State Space Model for Hyperspectral Image Reconstruction (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.