Research map / Image generation / Image restoration
HUNet: Homotopy Unfolding Network for Image 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 HUNet, a deep unfolding network that maps each homotopy-algorithm iteration to a multi-scale U-shaped Transformer module with learnable homotopy continuation and a dual-path feature fusion for image compressive sensing.
Problem
Existing deep unfolding networks for image compressive sensing underuse multi-scale features within each iteration and lose high-dimensional feature information between phases.
Approach
Unfold the homotopy algorithm into phases, each implemented by a U-shaped stack of window-based Transformer blocks with a deep (trainable, high-dimensional) homotopy continuation soft-thresholding step, and fuse inter-phase and intra-phase features at the end via a dual-path feature fusion module.
Key results
- Average PSNR on Urban100 at CS ratio 0.01-0.25 is 26.89 dB, versus 26.22 dB for the second-best method (NesTD-Net per table ordering), an average gain of about 0.64 dB.
- On Set11 at CS ratio 0.30/0.40/0.50, HUNet averages 39.80 dB PSNR / 0.9776 SSIM, the best among compared methods (CPP-Net 39.49 dB).
- Ablation: 7 phases chosen as default; the dual-path DFFM fusion outperforms single-path variants (Net-2, Net-3) on Set11, Set14, and Urban100.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the strongest reported baseline with HUNet | image compressive sensing reconstruction · LIVE29, BSD68, OST300, Set14, McM18, General100, Urban100 (dataset averages) | PSNR (dB): 27.16 -> 27.52; 27.24 -> 27.51; 26.77 -> 27.17; 28.14 -> 28.48; 30.95 -> 31.22; 30.75 -> 31.38; 26.25 -> 26.89 | Helped |
| replace CPP-Net with HUNet | image compressive sensing reconstruction · Set11 (average over CS ratios 0.30, 0.40, 0.50) | PSNR (dB): 39.49 -> 39.80 | Helped |
| increase the iterative phases from 3 to 7 (ablation) | HUNet · Set11 (CS ratio 0.04) | PSNR: 26.62 -> 27.23 | Helped |
| increase the iterative phases from 7 to 8 (ablation) | HUNet · Set11 (CS ratio 0.04) | PSNR: 27.23 -> 27.24 | No clear effect |
| add dual-path DFFM fusion of inter-phase and intra-phase features (ablation) | HUNet · Set11 (CS ratio 0.10) | PSNR: 30.87 -> 31.48 | Helped |
| add intra-phase features to inter-phase-only DFFM fusion (ablation) | HUNet · Set11 (CS ratio 0.10) | PSNR: 30.92 -> 31.48 | Helped |
| add inter-phase features to intra-phase-only DFFM fusion (ablation) | HUNet · Set11 (CS ratio 0.10) | PSNR: 31.23 -> 31.48 | Helped |
| add DHCS (ablation) | HUNet · Set11 (CS ratio 0.10) | PSNR: 31.35 -> 31.48 | Helped |
| increase iterative scale spaces from 1 to 3 (ablation) | HUNet · Set11 (CS ratio 0.10) | PSNR: 30.62 -> 31.48 | Helped |
| increase iterative scale spaces from 3 to 4 (ablation) | HUNet · Set11 (CS ratio 0.10) | PSNR: 31.48 -> 31.41 | No clear effect |
Limitations
- Parameter count and FLOPs are relatively high compared to some lightweight DUNs (e.g., UFC-Net)
- Gains over the strongest baselines are modest at high sampling rates in some datasets
- Evaluation limited to grayscale CS benchmarks with fixed patch-based sampling
Experimental setup
Datasets: Set11 LIVE29 BSD68 OST300 Set14 McM18 General100 Urban100
Benchmarks: LIVE29 BSD68 OST300 Set14 McM18 General100 Urban100 Set11
Metrics: PSNR SSIM
Baselines: DGU-Net+ TransCS OCTUF DPC-DUN TCS-Net CSformer LTwIST MTC-CSNet UFC-Net NesTD-Net DPA-Net CSNet+
Builds on: ISTA-Net homotopy algorithm OCTUF NesTD-Net CSformer
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)
- Using Powerful Prior Knowledge of Diffusion Model in Deep Unfolding Networks for Image Compressive Sensing (CVPR 2025 · similarity 0.88)
- UFC-Net: Unrolling Fixed-point Continuous Network for Deep Compressive Sensing (CVPR 2024 · similarity 0.87)
- Dual Prior Unfolding for Snapshot Compressive Imaging (CVPR 2024 · similarity 0.86)
- Untrained Neural Nets for Snapshot Compressive Imaging: Theory and Algorithms (NeurIPS 2024 · similarity 0.86)
- MCNet: Rethinking the Core Ingredients for Accurate and Efficient Homography Estimation (CVPR 2024 · 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.