Research map / Image generation / Image restoration
Sampling Innovation-Based Adaptive Compressive Sensing
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 SIB-ACS, an adaptive compressive sensing framework that allocates samples via a sampling-innovation criterion in a multi-stage negative-feedback loop, paired with a PCCD-Net reconstruction network.
Problem
Adaptive sampling allocation in compressive sensing of unknown scenes lacks accurate judgment and robust feedback, since ground truth is unavailable and existing measurement- or image-domain indicators are error-prone or form positive feedback loops.
Approach
Allocate samples to blocks by the sampling innovation: the predicted decrease in reconstruction error from an innovation sampling increment, computed from reconstructions before and after that increment. Iterating this over stages gives negative feedback, and a PCCD-Net unfolds PGD in both principal-component image and compressed feature domains to cut cost.
Key results
- SIB-ACS averages 35.83 dB on BSD68 and 36.31 dB on Urban100 across sampling rates 0.10-0.50, exceeding the second-best model by 1.10 dB and 1.48 dB respectively.
- Adaptive sampling beats uniform sampling with the same network: 29.54 vs 28.38 dB on BSD68 at SR=0.10.
- PCCD-Net with a compressed-domain dimension of 4 reaches 30.58 dB average PSNR at 2.32M parameters, near CPP-Net (30.63 dB at 12.47M).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace uniform sampling with innovation-based adaptive sampling (ablation) | PCCD-Net reconstruction · BSD68 and Urban100, SR 0.10 and 0.25 | PSNR (dB): BSD68 SR=0.10 28.38 to 29.54; Urban100 SR=0.25 33.11 to 35.15 | Helped |
| Use sampling innovation criterion instead of measurement-error or saliency to guide multi-stage adaptive sampling (ablation) | Multi-stage adaptive sampling with PCCD-Net · BSD68 and Urban100, SR 0.10 and 0.25 | PSNR (dB): Urban100 SR=0.25: 34.16 (measurement error) and 33.71 (saliency) vs 35.15 (innovation) | Helped |
| Add CDPGD path (remove PCPGD path vs full) (ablation) | PCCD-Net · BSD68 and Urban100, SR 0.10 and 0.25 | PSNR (dB): BSD68 SR=0.10 28.22 (no PCPGD) vs 28.38 full; Urban100 SR=0.10 28.10 vs 28.69 | Helped |
| Compressed domain dimension increased from 4 to 2, 8, 16, or 32 (ablation) | PCCD-Net · BSD68 and Urban100, SR 0.10 and 0.25 | PSNR (dB) and MACs: Quality oscillates with CD dimension; MACs rise from 2.2T (dim 4) to 13.8T (dim 32) at SR 0.25 | No clear effect |
| Use PCCD-Net (principal component compressed domain PGD) instead of CPP-Net-level heavy models | Image reconstruction · Average PSNR, BSD68 and Urban100 SR 0.10 and 0.25 | Params (M) and PSNR (dB): PCCD-Net 30.58 dB with 2.32M params vs CPP-Net 30.63 dB with 12.47M params | Mixed |
| Full SIB-ACS vs best uniform/adaptive baselines | SIB-ACS · BSD68 and Urban100 averages over SR 0.10-0.50 | PSNR (dB): +1.10 dB on BSD68 and +1.48 dB on Urban100 over second-best | Helped |
Surprising findings
- Saliency-guided multi-stage sampling shows positive feedback, concentrating samples in initially salient blocks.
- Increasing the compressed-domain dimension beyond 4 yields oscillating reconstruction quality while MACs grow.
Limitations
- Adaptive sampling evaluated only on BSD68 and Urban100 with block-based grayscale-style CS setup described in the text.
- Innovation estimation relies on lightweight reconstruction networks whose accuracy at low sampling rates is not analyzed in depth.
Experimental setup
Benchmarks: BSD68 Urban100
Metrics: PSNR SSIM
Baselines: AMP-Net-9-BM MADUN TransCS FSOINet OCTUF CPP-Net CASNet AMS-Net Uformer-ICS DGUNet+
Builds on: proximal gradient descent (PGD) deep unfolding networks
Closest papers by meaning
- Reconstruction-free Cascaded Adaptive Compressive Sensing (CVPR 2024 · similarity 0.94)
- Multi-Cross Sampling and Frequency-Division Reconstruction for Image Compressed Sensing (AAAI 2024 · similarity 0.85)
- Adapt and Diffuse: Sample-adaptive Reconstruction via Latent Diffusion Models (ICML 2024 · similarity 0.85)
- Provably Robust Score-Based Diffusion Posterior Sampling for Plug-and-Play Image Reconstruction (NeurIPS 2024 · similarity 0.84)
- Recovery Guarantees for Posterior Sampling of One-Bit Compressed Sensing (NeurIPS 2026 · similarity 0.84)
- DREAM: Diffusion Rectification and Estimation-Adaptive Models (CVPR 2024 · similarity 0.84)
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.