Research map / Architectures and efficiency / Attention and transformers
Dual Prior Unfolding for 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 Dual Prior Unfolding (DPU), combining an image prior and a degradation-based residual prior in a deep unfolding framework with Focused Attention and an asymmetric U-Net backbone for hyperspectral snapshot compressive imaging reconstruction.
Problem
Deep unfolding methods for snapshot compressive imaging (SCI) reconstruction use a single image prior, limiting iteration efficiency and making it hard to incorporate other priors such as degradation information; they also ignore shift and compression degradation.
Approach
A Dual Prior Framework adds a residual term with a degraded prior (learned from mask, shift and compression degradation) alongside the image prior, fused via gradient descent and residual learning; Focused Attention uses learnable principal-component scaling and threshold filtering of self-attention, plus an asymmetric U-Net backbone that shares hierarchy between local and shifted-window attention
Key results
- DPU-9stg reaches 40.52 dB PSNR and 0.977 SSIM averaged over 10 KAIST scenes, versus 39.62 dB for RDLUF.
- DPU-5stg outperforms RDLUF with less than 1/4 of its FLOPs, and gets 2.04 dB over BIRNAT with about 1/77 of its FLOPs.
- In the framework comparison, the DPF gives 1.23 dB over GAP and 0.78 dB over RDLF at 5 stages.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DPU-9stg instead of RDLUF | SCI reconstruction · KAIST 10-scene simulation, Avg | PSNR: 39.57 -> 40.52 | Helped |
| Use DPU-9stg instead of RDLUF | SCI reconstruction · KAIST 10-scene simulation, Avg | SSIM: 0.974 -> 0.977 | Helped |
| Add L/Swin-FA (ablation) | Baseline-1 · KAIST 10-scene simulation, Avg | PSNR: 37.28 -> 38.49 | Helped |
| Add the intuitive DPF framework (ablation) | Baseline-1 + L/Swin-FA · KAIST 10-scene simulation, Avg | PSNR: 38.49 -> 38.76 | Helped |
| Add the basic DPF framework (ablation) | Baseline-1 + L/Swin-FA · KAIST 10-scene simulation, Avg | PSNR: 38.49 -> 39.23 | Helped |
| Add the complete DPF framework (ablation) | Baseline-1 + L/Swin-FA · KAIST 10-scene simulation, Avg | PSNR: 38.49 -> 39.62 | Helped |
| Use the asymmetric backbone with Swin-MSA (ablation) | Baseline-2 attention module · KAIST 10-scene simulation, Avg | PSNR: 35.97 -> 36.27 | Helped |
| Add focused attention to Swin* (ablation) | Swin* · KAIST 10-scene simulation, Avg | PSNR: 36.27 -> 36.58 | Helped |
Surprising findings
- The 5-stage DPF outperforms other frameworks' 9-stage variants while using less computation.
Limitations
- Evaluated only on hyperspectral SCI reconstruction with CASSI-style setups
- Ablations are reported on the 5-stage model only
Experimental setup
Datasets: CAVE KAIST
Benchmarks: KAIST 10 simulation scenes real CASSI measurements (TSA-Net dataset)
Metrics: PSNR SSIM FLOPs Params
Baselines: TSA-Net DGSMP GAP-Net HDNet MST CST BIRNAT DAUHST RDLUF
Builds on: DAUHST MST RDLUF Swin Transformer GAP-Net
Closest papers by meaning
- Untrained Neural Nets for Snapshot Compressive Imaging: Theory and Algorithms (NeurIPS 2024 · similarity 0.87)
- Using Powerful Prior Knowledge of Diffusion Model in Deep Unfolding Networks for Image Compressive Sensing (CVPR 2025 · similarity 0.87)
- CPP-Net: Embracing Multi-Scale Feature Fusion into Deep Unfolding CP-PPA Network for Compressive Sensing (CVPR 2024 · similarity 0.87)
- A Principled Self-Referenced Early Stopping Approach for Deep Image Prior (NeurIPS 2026 · similarity 0.86)
- HUNet: Homotopy Unfolding Network for Image Compressive Sensing (CVPR 2025 · similarity 0.86)
- Provably Robust Score-Based Diffusion Posterior Sampling for Plug-and-Play Image Reconstruction (NeurIPS 2024 · 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.