Research map / 3D vision / Depth and stereo
Reconstruction-free Cascaded 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 a reconstruction-free cascaded adaptive compressive sensing framework that allocates multi-stage sampling via a lightweight ScoreNet and LP-based differentiable sampling, with a multi-grid spatial-attention reconstruction network (MGSANet).
Problem
Multi-stage adaptive compressive sensing requires image reconstruction within the sampling loop to steer successive measurements, which slows sampling and hinders time-sensitive applications.
Approach
Score blocks directly from previous CS measurements with a lightweight ScoreNet, allocate sampling rates via a perturbed-optimizer LP-based differentiable selection mask, and supervise reconstruction fidelity out-of-loop at each stage so that only one final reconstruction is needed.
Key results
- On Set11 the method reaches 34.10 dB average PSNR versus 33.96 for AMS-Net and 33.59 for OCTUF+ across 10-25% sampling rates.
- On Urban100 the method averages 31.86 dB PSNR/0.9286 SSIM, outperforming OCTUF+ (30.66 dB) by about 1.2 dB.
- MGSANet reconstructs roughly 6x faster than OCTUF+ on Urban100 at 25% sampling rate (per the paper's claim).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use adaptive sampling instead of uniform sampling (ablation) | MGSANet · Set11, 10% sampling | PSNR: 30.24 -> 31.05 | Helped |
| use adaptive sampling instead of uniform sampling (ablation) | MGSANet · Urban100, 10% sampling | PSNR: 28.10 -> 29.09 | Helped |
| include the perturbed optimizer in differentiable ASM (ablation) | Reconstruction-free ACS · Set11, 10% sampling | PSNR: 30.38 -> 31.05 | Helped |
| include the perturbed optimizer in differentiable ASM (ablation) | Reconstruction-free ACS · Urban100, 10% sampling | PSNR: 28.33 -> 29.09 | Helped |
| apply reconstruction-fidelity supervision at each stage rather than only at the last stage (ablation) | Reconstruction-free ACS · Set11, 25% sampling | PSNR: 35.65 -> 36.62 | Helped |
| apply reconstruction-fidelity supervision at each stage rather than only at the last stage (ablation) | Reconstruction-free ACS · Urban100, 25% sampling | PSNR: 33.23 -> 34.27 | Helped |
| use spatial attention rather than CBAM (ablation) | MGSANet · Urban100, 25% sampling | PSNR: 32.57 -> 32.92 | Helped |
| use 4 scales rather than 2 scales (ablation) | MGSANet · Set11, 25% sampling | PSNR: 34.43 -> 35.35 | Helped |
| use 5 scales rather than 4 scales (ablation) | MGSANet · Urban100, 25% sampling | PSNR: 32.91 -> 32.76 | Hurt |
Surprising findings
- Adaptive sampling without reconstruction in the loop still beats the best non-adaptive methods, while AMS-Net uses ground-truth access for its sampling scheme.
Limitations
- MGSANet requires more FLOPs (1.55T on Urban100 at 25%) than several baselines.
- Training is stage-by-stage and requires multiple training phases per stage.
- Evaluation limited to grayscale Y channel on Set11 and Urban100.
Experimental setup
Datasets: BSDS500 VOC2012
Benchmarks: Set11 Urban100
Metrics: PSNR SSIM
Baselines: AMP-Net OPINE-Net+ COAST NL-CS MADUN TransCS FSOINet CSFormer TCS-Net OCTUF+ ACCSNet CASNet
Builds on: perturbed optimizer ConvNeXt CASNet ACCSNet
Closest papers by meaning
- Sampling Innovation-Based Adaptive Compressive Sensing (CVPR 2025 · similarity 0.94)
- Multi-Cross Sampling and Frequency-Division Reconstruction for Image Compressed Sensing (AAAI 2024 · similarity 0.86)
- UFC-Net: Unrolling Fixed-point Continuous Network for Deep Compressive Sensing (CVPR 2024 · similarity 0.86)
- CPP-Net: Embracing Multi-Scale Feature Fusion into Deep Unfolding CP-PPA Network for Compressive Sensing (CVPR 2024 · similarity 0.85)
- Adapt and Diffuse: Sample-adaptive Reconstruction via Latent Diffusion Models (ICML 2024 · similarity 0.84)
- Untrained Neural Nets for Snapshot Compressive Imaging: Theory and Algorithms (NeurIPS 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.