Research map / Architectures and efficiency / CNN design
Block Image Compressive Sensing with Local and Global Information Interaction
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 BRBCN, a deep unfolding block compressive sensing network that combines a CNN (per-block reconstruction) and a Transformer (inter-block communication) with bidirectional G2L/L2G modules, plus a BRBCN+ variant adding an image-level CNN.
Problem
Block-based image compressive sensing reconstructs each block independently, ignoring global communication among blocks, while existing fixes (local filters or image-level reconstruction) either provide insufficient inter-block communication or lose inner-block detail.
Approach
Dual CNN and Transformer branches inside a deep unfolding framework: the CNN (UNet) reconstructs each block locally while ViT tokens model inter-block communication, with G2L and L2G modules passing information both ways at every stage.
Key results
- BRBCN improves average PSNR over the second-best method by 0.36 dB on Set14 and 0.21 dB on BSD68 across sampling ratios (grayscale).
- BRBCN+ reaches 30.44 dB average PSNR on Set14 and 31.86 dB on Waterloo, the best reported averages.
- Adding global token communication without fusion lifts Set14 PSNR at ratio 0.01 from 20.67 to 21.87 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add block-level global inter-block communication (ablation) | BRBCN · Set14, ratio 0.01 | PSNR: 20.67 -> 21.87 | Helped |
| add block-level global inter-block communication (ablation) | BRBCN · BSD68, ratio 0.01 | PSNR: 21.46 -> 22.54 | Helped |
| add image-level IBR | BRBCN · Set14, sampling ratio 0.5 | PSNR: 39.56 -> 39.44 | Hurt |
| add image-level IBR | BRBCN · Set14, sampling ratio 0.01 | PSNR: 21.87 -> 22.39 | Helped |
| use serial CNN-Transformer fusion instead of fusion-free CNN (ablation) | BRBCN · Set14, ratio 0.01 | PSNR: 21.42 -> 20.45 | Hurt |
| use parallel CNN-Transformer fusion instead of fusion-free CNN (ablation) | BRBCN · Set14, ratio 0.01 | PSNR: 21.42 -> 21.56 | Helped |
| replace G2L/L2G with Mobile-Former fusion (ablation) | BRBCN · Set14, ratio 0.01 | PSNR: 21.42 -> 20.23 | Hurt |
| use proposed G2L and L2G fusion (ablation) | BRBCN · Set14, ratio 0.01 | PSNR: 21.42 -> 21.87 | Helped |
| add L2G local-to-global interaction alongside G2L (ablation) | BRBCN · Set14, ratio 0.1 | PSNR: 28.65 -> 29.51 | Helped |
| add L2G local-to-global interaction alongside G2L (ablation) | BRBCN · BSD68, ratio 0.1 | PSNR: 27.74 -> 28.22 | Helped |
| increase iteration count K (ablation) | BRBCN · Set14, ratio 0.1 | PSNR | No clear effect |
| increase token dimension CT beyond 128 (ablation) | BRBCN · Set14, ratio 0.1 | PSNR | No clear effect |
Surprising findings
- Simple serial or parallel fusion of CNN and Transformer features performs worse than CNN-only, and Mobile-Former fusion performs worse than simple parallel combination.
- BRBCN+ underperforms BRBCN at the 0.5 sampling ratio because the image-level IBR harms fine-grained block details.
Limitations
- High parameter count and computational cost driven by the maximum channel setting in IBR (up to 43M params for BRBCN 256).
- Robustness to noisy inputs and generalization to multi-scale sampling are not yet addressed.
- BRBCN+ is weaker than BRBCN at 0.5 sampling ratio due to the image-level IBR.
Experimental setup
Datasets: ImageNet Set14 BSD68 Waterloo
Benchmarks: Set14 BSD68 Waterloo
Metrics: PSNR SSIM running time parameter count
Baselines: ISTA-Net+ CSNet+ AMP-Net OPINE-Net+ MADUN DGUNet+
Builds on: ISTA-Net ADMM-CSNet Mobile-Former UNet Vision Transformer
Closest papers by meaning
- Untrained Neural Nets for Snapshot Compressive Imaging: Theory and Algorithms (NeurIPS 2024 · similarity 0.85)
- UFC-Net: Unrolling Fixed-point Continuous Network for Deep Compressive Sensing (CVPR 2024 · similarity 0.84)
- Multi-Cross Sampling and Frequency-Division Reconstruction for Image Compressed Sensing (AAAI 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)
- HUNet: Homotopy Unfolding Network for Image Compressive Sensing (CVPR 2025 · similarity 0.83)
- Using Powerful Prior Knowledge of Diffusion Model in Deep Unfolding Networks for Image Compressive Sensing (CVPR 2025 · 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.