Research map / Architectures and efficiency / Attention and transformers
Dual-Scale Transformer for Large-Scale Single-Pixel 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 HATNet, a deep unfolding network with hybrid-attention Transformer denoisers on Kronecker single-pixel imaging, enabling full-size sampling and reconstruction of large images on a real SPI prototype.
Problem
Deep unfolding networks for single-pixel imaging rely on block-based sampling or vectorized models that are impractical for real SPI cameras and computationally expensive at large resolutions.
Approach
Unfold tensor ISTA on the Kronecker SPI model so gradient descent uses two small measurement matrices, and use a U-shaped denoiser with dual-scale spatial self-attention (high/low-frequency branches via average pooling) plus channel self-attention.
Key results
- On Set11 at 10% sampling ratio, HATNet reaches 32.26 dB / 0.9182, above SAUNet's 32.15 / 0.9147.
- Kronecker SPI reduces GPU memory from 10.74 G to 3.02 G and inference time from 0.55 s to 0.38 s versus vectorized SPI.
- On a real prototype at 768x1024 resolution and 12.5% sampling ratio, HATNet outperforms ISTA-TV by a large margin (no numbers given).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use HATNet instead of the strongest reported baseline, SAUNet | SPI reconstruction · Set11, 4% SR | PSNR: 27.80 -> 27.98 | Helped |
| use HATNet instead of the strongest reported baseline, SAUNet | SPI reconstruction · Set11, 10% SR | PSNR: 32.15 -> 32.26 | Helped |
| use HATNet instead of the strongest reported baseline, SAUNet | SPI reconstruction · Set11, 25% SR | PSNR: 37.11 -> 37.24 | Helped |
| use HATNet instead of the strongest reported baseline, SAUNet | SPI reconstruction · Set11, 50% SR | PSNR: 41.91 -> 42.05 | Helped |
| add cross-stage skip connections (CSSC) (ablation) | HATNet · Set11, 10% SR | PSNR: 31.48 -> 32.26; 0.78 dB | Helped |
| use dual-scale high- and low-frequency aggregation in S-SA rather than LF-only aggregation (ablation) | HATNet · Set11, 10% SR | PSNR: 31.56 -> 32.26; 0.70 dB | Helped |
| use dual-scale high- and low-frequency aggregation in S-SA rather than HF-only aggregation (ablation) | HATNet · Set11, 10% SR | PSNR: 31.78 -> 32.26; 0.48 dB | Helped |
| add channel-wise self-attention (C-SA) (ablation) | HATNet · Set11, 10% SR | PSNR: 32.15 -> 32.26; 0.11 dB | Helped |
| use full-size Kronecker SPI instead of block-based sampling (ablation) | HATNet · Set11, 10% SR | PSNR: 31.62 -> 32.26 | Helped |
| use Kronecker SPI instead of vectorized SPI (ablation) | HATNet · Set11, 25% SR | GPU memory: 10.74 -> 3.02 | Helped |
Surprising findings
- Re-training block-based DUNs (DGUNet+, OCTUF+) under Kronecker full-size sampling improves them, but re-training HATNet under block-based sampling lowers its performance.
Limitations
- Large-scale real results are only qualitative; no quantitative comparison against other DUNs due to training cost.
- Requires learnable or Hadamard-based dual measurement matrices, which real hardware must support (cake-cutting Hadamard).
Experimental setup
Datasets: BSD500 DIV2K Set11
Benchmarks: Set11 real SPI prototype captures
Metrics: PSNR SSIM
Baselines: ReconNet ISTA-Net+ CSNet+ SCSNet OPINE-Net+ AMP-Net TransCS MADUN DGUNet+ OCTUF+ SAUNet ISTA-TV
Builds on: ISTA tensor ISTA unfolding (SAUNet) Restormer PVT DGUNet
Closest papers by meaning
- Proximal Algorithm Unrolling: Flexible and Efficient Reconstruction Networks for Single-Pixel Imaging (CVPR 2025 · similarity 0.85)
- Progressive Focused Transformer for Single Image Super-Resolution (CVPR 2025 · similarity 0.84)
- Recursive Generalization Transformer for Image Super-Resolution (ICLR 2024 · similarity 0.84)
- CPP-Net: Embracing Multi-Scale Feature Fusion into Deep Unfolding CP-PPA Network for Compressive Sensing (CVPR 2024 · similarity 0.84)
- Attention-Only Transformers via Unrolled Subspace Denoising (ICML 2025 · similarity 0.84)
- Dual-Window Multiscale Transformer for Hyperspectral Snapshot Compressive Imaging (AAAI 2024 · similarity 0.83)
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.