Research map / Architectures and efficiency / Attention and transformers
UFC-Net: Unrolling Fixed-point Continuous Network for Deep 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 UFC-Net, a deep unfolding compressive sensing network based on the fixed-point continuation (FPC) algorithm, with a convolution-guided attention module (MARB + AIRB) for image CS and CS-MRI.
Problem
Deep unfolding compressive sensing networks suffer from insufficient feature extraction (local CNN bias, limited Transformer local modeling) and feature loss between iterative stages when reconstructing images from undersampled measurements.
Approach
Unrolls the fixed-point continuation algorithm into stages with learnable gradient descent steps and a Convolution-guided Attention Module that combines multi-head self-attention with convolutions (MARB) and auxiliary variables passed between stages (AIRB) to preserve inter-stage information.
Key results
- On CIFAR10 at sampling rate 0.01, UFC-Net reaches 18.29 dB PSNR, 0.68 dB above OCTUF's 17.61 dB (per the paper's stated gain relative to the prior table entries).
- On Urban100 at sampling rate 0.10, UFC-Net achieves 27.55 dB PSNR and 0.8583 SSIM, improving on the second-best method by 0.61 dB.
- On FastMRI at sampling rate 0.15, UFC-Net achieves 29.46 dB PSNR, 0.36 dB above HiTDUN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UFC-Net instead of OCTUF | Image compressive sensing · CIFAR10 and CIFAR100, δ=0.10 | PSNR: 25.26 -> 25.39; 25.86 -> 25.96 | Helped |
| Use UFC-Net instead of the strongest listed baseline | Image compressive sensing · Set11, Set14, Urban100, and General100, δ=0.10 | PSNR: 29.84 -> 30.15; 28.82 -> 29.10; 26.94 -> 27.55; 31.91 -> 32.31 | Helped |
| Use UFC-Net instead of HiTDUN | CS-MRI · FastMRI, δ=0.15 | PSNR: 29.10 -> 29.46 | Helped |
| Add AIRB (ablation) | UFC-Net · Set5, Set11, and McM18, δ=0.25 | PSNR: 36.88 -> 37.08; 35.16 -> 35.42; 37.03 -> 37.24 | Helped |
| Add MARB (ablation) | UFC-Net · Set5, Set11, and McM18, δ=0.25 | PSNR: 36.92 -> 37.08; 35.11 -> 35.42; 37.05 -> 37.24 | Helped |
Limitations
- Parameter count and GFLOPs are described by the authors as not optimal compared with some baselines.
- Feature extraction and inter-stage loss issues are addressed only through architectural changes; no analysis beyond the reported ablations.
Experimental setup
Datasets: CIFAR10 CIFAR100 FastMRI knee
Benchmarks: CIFAR10 CIFAR100 Set11 Set14 Urban100 General100 Set5 McM18 FastMRI
Metrics: PSNR SSIM parameters GFLOPs GPU running time
Baselines: ISTA-Net+ CSGAN CSNet+ DPA-Net OPINE-Net+ MAC-Net NL-CSNet AMP-Net COAST FSOINET TransCS DGUNet+
Builds on: ISTA-Net fixed-point continuation ConvNeXt V2 AMP-Net
Closest papers by meaning
- CPP-Net: Embracing Multi-Scale Feature Fusion into Deep Unfolding CP-PPA Network for Compressive Sensing (CVPR 2024 · similarity 0.89)
- HUNet: Homotopy Unfolding Network for Image Compressive Sensing (CVPR 2025 · similarity 0.87)
- Using Powerful Prior Knowledge of Diffusion Model in Deep Unfolding Networks for Image Compressive Sensing (CVPR 2025 · similarity 0.87)
- Untrained Neural Nets for Snapshot Compressive Imaging: Theory and Algorithms (NeurIPS 2024 · similarity 0.86)
- A Unified Model for Compressed Sensing MRI Across Undersampling Patterns (CVPR 2025 · similarity 0.86)
- Multi-Cross Sampling and Frequency-Division Reconstruction for Image Compressed Sensing (AAAI 2024 · similarity 0.86)
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.