Research map / Image generation / Image restoration
Multi-Cross Sampling and Frequency-Division Reconstruction for Image Compressed 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 MCFD-Net, a deep compressed sensing network whose DMCS sampling module uses pyramid cross convolution with dual-branch pooling and attention, and whose FDRM reconstruction module splits features via DWT into high- and low-frequency branches for separate multi-scale convolution and self-attention compensation.
Problem
Deep compressed sensing methods use simple convolutional downsampling for measurement, losing high-level signal features, and ignore separate handling of high- and low-frequency information during reconstruction.
Approach
A learned sampling network (DMCS) with pyramid cross convolution, dual-branch multi-level pooling and attention replaces simple strided convolution for measurement, and a reconstruction module (FDRM) applies a discrete wavelet transform to separate frequency bands, processing each with multi-scale convolution and self-attention before fusion.
Key results
- On Set11 at 50% sampling, MCFD-Net reaches 42.92 dB PSNR, versus 42.37 dB for OCTUF and 37.67 dB for RK-CCSNet.
- On BSDS100 at 25% sampling, MCFD-Net achieves 35.71 dB PSNR, 1.48 dB above TransCS and 0.92 dB above OCTUF.
- Under Gaussian noise (variance up to 0.020) on Urban100 at 25% sampling, MCFD-Net degrades by only 1.05 dB, versus 2.52 dB for DPC-DUN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the strongest listed baseline with MCFD-Net | Image compressed sensing · Set11, 50% sampling | PSNR (dB): 42.37 -> 42.92 | Helped |
| Replace the strongest listed baseline with MCFD-Net | Image compressed sensing · Set11, 25% sampling | PSNR (dB): 35.67 -> 36.72 | Helped |
| Replace the strongest listed baseline with MCFD-Net | Image compressed sensing · BSDS100, 25% sampling | PSNR (dB): 34.79 -> 35.71 | Helped |
| Add FDRM to the MR-CCSNet+ configuration (ablation) | MCFD-Net on BSDS100 · 25% sampling | PSNR (dB): 33.96 -> 34.37 | Helped |
| Add FDRM to the MR-CCSNet+ configuration (ablation) | MCFD-Net on BSDS100 · 1.5625% sampling | PSNR (dB): 23.90 -> 24.26 | Helped |
| Add DMCS to the MR-CCSNet+ configuration (ablation) | MCFD-Net on BSDS100 · 25% sampling | PSNR (dB): 33.96 -> 35.40 | Helped |
| Add DMCS to the MR-CCSNet+ configuration (ablation) | MCFD-Net on BSDS100 · 1.5625% sampling | PSNR (dB): 23.90 -> 24.34 | Helped |
| Add both DMCS and FDRM (ablation) | MCFD-Net on BSDS100 · 25% sampling | PSNR (dB): 33.96 -> 35.71 | Helped |
| Add both DMCS and FDRM (ablation) | MCFD-Net on BSDS100 · 1.5625% sampling | PSNR (dB): 23.90 -> 24.65 | Helped |
| Use DMCS instead of FDRM alone (ablation) | MCFD-Net on BSDS100 · 25% sampling | PSNR (dB): 34.37 -> 35.40 | Helped |
| Use DMCS instead of FDRM alone (ablation) | MCFD-Net on BSDS100 · 1.5625% sampling | PSNR (dB): 24.26 -> 24.34 | Helped |
| Use MCFD-Net rather than DPC-DUN under Gaussian noise | Urban100 at 25% sampling · Gaussian noise variance 0.020 | PSNR (dB): 28.22 -> 31.87 | Helped |
Surprising findings
- Ablation shows the DMCS sampling module contributes more to performance than the FDRM reconstruction module, suggesting improved sampling yields larger gains than improved reconstruction.
Limitations
- At 12.5% and lower sampling rates, MCFD-Net is slower than DPC-DUN and OCTUF on the reported CPU/GPU timing.
- Evaluation is limited to grayscale Y channel on three benchmark datasets.
Experimental setup
Datasets: COCO 2014
Benchmarks: Set11 BSDS100 Urban100
Metrics: PSNR SSIM
Baselines: CSNet+ RK-CCSNet NL-CSNet MR-CCSNet+ TransCS CASNet DPC-DUN OCTUF
Builds on: MR-CCSNet CSNet+ NL-CSNet residual learning
Closest papers by meaning
- CPP-Net: Embracing Multi-Scale Feature Fusion into Deep Unfolding CP-PPA Network for Compressive Sensing (CVPR 2024 · similarity 0.87)
- Using Powerful Prior Knowledge of Diffusion Model in Deep Unfolding Networks for Image Compressive Sensing (CVPR 2025 · similarity 0.86)
- Reconstruction-free Cascaded Adaptive Compressive Sensing (CVPR 2024 · similarity 0.86)
- UFC-Net: Unrolling Fixed-point Continuous Network for Deep Compressive Sensing (CVPR 2024 · similarity 0.86)
- Untrained Neural Nets for Snapshot Compressive Imaging: Theory and Algorithms (NeurIPS 2024 · similarity 0.86)
- Sampling Innovation-Based Adaptive Compressive Sensing (CVPR 2025 · similarity 0.85)
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.