Research map / Image generation / Image restoration
CPP-Net: Embracing Multi-Scale Feature Fusion into Deep Unfolding CP-PPA Network for 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 CPP-Net, a deep unfolding compressive sensing network derived from the Chambolle-Pock proximal point algorithm, adding a multi-scale Dual Path Fusion Block and an Iteration Fusion Strategy to better recover fine image details at low CS ratios.
Problem
Deep unfolding networks for compressive sensing image reconstruction produce distortions and blurring, especially at low CS ratios (0.10 and below), because they fail to extract and fuse multi-scale features and to fuse information across reconstruction stages.
Approach
Unfolds the primal-dual CP-PPA into three learnable iteration submodules (X, V, Y) with a learned sparse-basis operator, and augments each stage with a multi-scale Dual Path Fusion Block and a final attention-weighted Iteration Fusion Strategy over all stage outputs.
Key results
- At CS ratio 0.10 on Set14, CPP-Net reaches 29.93 dB / 0.8537 SSIM, versus 29.47 dB for OCTUF and 29.34 dB for DGUNet+.
- On General100 at CS ratio 0.04, CPP-Net surpasses OCTUF by about 0.69 dB in PSNR.
- On Brain CS-MRI at CS ratio 0.05, CPP-Net reaches 34.87 dB / 0.9176 SSIM, versus 32.72 dB for HiTDUN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the strongest reported baseline with CPP-Net | compressive sensing reconstruction · General100 Avg. | PSNR: 30.71 -> 31.17 | Helped |
| replace the strongest reported baseline with CPP-Net | compressive sensing reconstruction · LIVE29 Avg. | PSNR: 27.15 -> 27.35 | Helped |
| replace the strongest reported baseline with CPP-Net | compressive sensing reconstruction · McM18 Avg. | PSNR: 30.74 -> 31.15 | Helped |
| replace the strongest reported baseline with CPP-Net | compressive sensing reconstruction · OST300 Avg. | PSNR: 26.76 -> 26.94 | Helped |
| replace the strongest reported baseline with CPP-Net | compressive sensing reconstruction · Set11 Avg. | PSNR: 29.02 -> 29.26 | Helped |
| replace the strongest reported baseline with CPP-Net | compressive sensing reconstruction · Set14 Avg. | PSNR: 27.91 -> 28.35 | Helped |
| replace the strongest reported baseline with CPP-Net | compressive sensing reconstruction · Urban100 Avg. | PSNR: 26.25 -> 26.77 | Helped |
| replace the strongest reported baseline with CPP-Net | CS-MRI reconstruction · Brain MRI Avg., Pseudo Radial masks | PSNR: 35.90 -> 37.05 | Helped |
| increase reconstruction stages from 3 to 8 (ablation) | CPP-Net · Set14, CS ratio 0.25 | PSNR: 33.65 -> 34.41 | Helped |
| increase reconstruction stages from 8 to 9 (ablation) | CPP-Net · Set14, CS ratio 0.25 | PSNR: 34.41 -> 34.46 | Helped |
| add STRes, downsampling, upsampling, DCAB, residual blocks, and MSTB (ablation) | Net-1 · Set11, CS ratio 0.10 | PSNR: 30.15 -> 31.24 | Helped |
| remove IFS and use shared reconstruction-stage parameters (ablation) | CPP-Net · Set11, CS ratio 0.10 | PSNR: 31.27 -> 30.75 | Hurt |
Limitations
- Parameter count (12.31M) and FLOPs (153.47G) are higher than several baselines such as OCTUF and FSOINet.
- Evaluated only on natural image CS and brain MRI; generalization to other restoration tasks is left to future work.
Experimental setup
Datasets: Set11 training images (implied by ISTA-Net style training) Brain MRI dataset (CS-MRI)
Benchmarks: General100 LIVE29 McM18 OST300 Set11 Set14 Urban100 Brain MRI (Pseudo Radial masks)
Metrics: PSNR SSIM LPIPS
Baselines: ISTA-Net+ AMP-Net CASNet DGUNet+ FSOINet TransCS AutoBCS CSformer DPC-DUN OCTUF TCS-Net DC-CNN
Builds on: CP-PPA ISTA-Net DGUNet+ OCTUF ADMM-CSNet
Closest papers by meaning
- HUNet: Homotopy Unfolding Network for Image Compressive Sensing (CVPR 2025 · similarity 0.89)
- UFC-Net: Unrolling Fixed-point Continuous Network for Deep Compressive Sensing (CVPR 2024 · similarity 0.89)
- Using Powerful Prior Knowledge of Diffusion Model in Deep Unfolding Networks for Image Compressive Sensing (CVPR 2025 · similarity 0.89)
- Multi-Cross Sampling and Frequency-Division Reconstruction for Image Compressed Sensing (AAAI 2024 · similarity 0.87)
- Dual Prior Unfolding for Snapshot Compressive Imaging (CVPR 2024 · similarity 0.87)
- Untrained Neural Nets for Snapshot Compressive Imaging: Theory and Algorithms (NeurIPS 2024 · similarity 0.87)
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.