Research map / Architectures and efficiency / Attention and transformers
Detail Matters: Mamba-Inspired Joint Unfolding Network for Snapshot Spectral 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 MiJUN, a deep unfolding network for snapshot spectral compressive imaging that combines an accelerated HQS scheme with a Mamba-inspired global-local attention block and mode-k tensor unfolding scanning.
Problem
Hyperspectral image reconstruction from a single coded-aperture snapshot (CASSI) measurement is ill-posed and nonlinear; existing deep unfolding networks trade off accuracy, stability, computational cost, and long-range versus local detail modeling.
Approach
Combines an accelerated half-quadratic splitting (A-HQS) unfolding scheme with a Mamba block reinterpreted as a Transformer-like variant (adding a global-local attention branch), and applies tensor mode-k unfolding before the SSM to enable 12-direction scanning emphasizing low-rank structure.
Key results
- MiJUN-9stg reaches 40.86 dB average PSNR and 0.982 SSIM on 10 simulated KAIST scenes, versus 39.57 dB for RDLUF-MixS2-9stg.
- MiJUN-5stg reaches 40.60 dB average PSNR with 0.56M parameters and 40.98 GFLOPs, versus 1.89M parameters for RDLUF-MixS2-9stg.
- Ablation shows Mamba-i and M-k Mamba add about 1.3 dB and 0.35 dB PSNR over the baseline and prior Mamba variant respectively.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add acceleration strategy (ablation) | RDLUF-MixS²-5stg ablation baseline · Simulated HSI ablation | PSNR (dB): 38.59 -> 38.60 | Helped |
| Add acceleration strategy (ablation) | RDLUF-MixS²-5stg ablation baseline · Simulated HSI ablation | SSIM: 0.969 -> 0.971 | Helped |
| Add acceleration and replace attention with GLAM (ablation) | RDLUF-MixS²-5stg ablation baseline · Simulated HSI ablation | PSNR (dB): 38.59 -> 39.64 | Helped |
| Add acceleration and replace attention with GLAM (ablation) | RDLUF-MixS²-5stg ablation baseline · Simulated HSI ablation | SSIM: 0.969 -> 0.974 | Helped |
| Add Mamba-i T (ablation) | MiJUN ablation model with GLAM · Simulated HSI ablation | PSNR (dB): 39.64 -> 39.89 | Helped |
| Add Mamba-i T (ablation) | RDLUF-MixS²-5stg ablation baseline · Simulated HSI ablation | PSNR (dB): +1.3 dB | Helped |
| Add Mamba module (ablation) | MiJUN ablation model · Simulated HSI ablation | PSNR (dB): 39.89 -> 40.25 | Helped |
| Add mode-k tensor unfolding to Mamba (ablation) | MiJUN ablation model · Simulated HSI ablation | PSNR (dB): 40.25 -> 40.60 | Helped |
| Add mode-k tensor unfolding to Mamba (ablation) | MiJUN ablation model · Simulated HSI ablation | SSIM: 0.976 -> 0.978 | Helped |
| Use MiJUN instead of RDLUF | ICVL unseen test · eve_0311 | PSNR: 31.29 -> 31.97 | Helped |
| Use MiJUN instead of RDLUF | ICVL unseen test · BUG-0403 | PSNR: 30.74 -> 31.27 | Helped |
| Use MiJUN instead of RDLUF | ICVL unseen test · 4cam_0411 | PSNR: 32.90 -> 33.82 | Helped |
Limitations
- Real-data results still show blurriness and artifacts, per the authors.
- Ablation and comparisons are reported on a single simulated benchmark setting, with real data evaluated only qualitatively.
Experimental setup
Datasets: CAVE KAIST ICVL five real CASSI datasets
Benchmarks: KAIST simulation (10 scenes) ICVL generalization real CASSI datasets
Metrics: PSNR SSIM
Baselines: TwIST GAP-TV DeSCI HDNet BIRNAT DADF-Plus-3 MST++ CST-L+ DNU GAP-Net DAUHST-9stg PADUT-5stg
Builds on: Mamba Vision Mamba half-quadratic splitting DAUHST RDLUF-MixS2
Closest papers by meaning
- Sp3ctralMamba: Physics-Driven Joint State Space Model for Hyperspectral Image Reconstruction (AAAI 2025 · similarity 0.89)
- Spectral Compressive Imaging via Chromaticity-Intensity Decomposition (NeurIPS 2025 · similarity 0.86)
- MambaSCI: Efficient Mamba-UNet for Quad-Bayer Patterned Video Snapshot Compressive Imaging (NeurIPS 2024 · similarity 0.86)
- Untrained Neural Nets for Snapshot Compressive Imaging: Theory and Algorithms (NeurIPS 2024 · similarity 0.86)
- Dual-Window Multiscale Transformer for Hyperspectral Snapshot Compressive Imaging (AAAI 2024 · similarity 0.86)
- SSUN-Net: Spatial-Spectral Prior-Aware Unfolding Network for Pan-Sharpening (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.