Research map / Image generation / Super-resolution
OSFFNet: Omni-Stage Feature Fusion Network for Lightweight Image Super-Resolution
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 OSFFNet, a lightweight single-image super-resolution network that fuses multi-stage shallow and deep features via stacked original-image initialization, shallow global connections, multi-receptive-field dynamic fusion, and attention-enhanced feature distillation.
Problem
Lightweight single-image super-resolution underuses shallow features (edges, textures, colors) because sequentially simplified networks lose low-level detail during forward propagation.
Approach
An Omni-Stage Feature Fusion architecture that stacks the original LR image along channels for shallow feature extraction, adds a learnable-weighted shallow-feature global connection to every block, and fuses per-stage features using dynamic receptive-field modules with varying kernel sizes, together with an attention-enhanced (ESA+CCA) feature distillation block built on BSConv.
Key results
- OSFFNet achieves 0.26 dB PSNR gain over the second-best method for x2 SR on Urban100 (per abstract).
- For x4 SR on Urban100, OSFFNet reports 26.36 dB PSNR (SSIM 0.7950) with 537K params and 22.0G Multi-Adds, the best across listed datasets at x4.
- Applying the OSFF architecture to IMDN improves x4 Urban100 PSNR from 26.04 to 26.11 dB with a modest parameter increase.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use OSFFNet instead of the strongest listed baseline | Lightweight SISR · Urban100 ×2 | PSNR: 32.34 -> 32.67 | Helped |
| Use OSFFNet instead of the strongest listed baseline | Lightweight SISR · Urban100 ×4 | PSNR: 26.28 -> 26.36 | Helped |
| Use OSFFNet instead of the strongest listed baseline | Lightweight SISR · Set14 ×2 | PSNR: 33.75 -> 33.72 | Hurt |
| Use OSFFNet instead of the strongest listed baseline | Lightweight SISR · Manga109 ×2 | PSNR: 39.14 -> 39.09 | Hurt |
| Add OISI to the configuration without OSFF modules (ablation) | OSFFNet · Urban100 ×4 | PSNR: 25.76 -> 25.98 | Helped |
| Add SFGC to the configuration without OSFF modules (ablation) | OSFFNet · Urban100 ×4 | PSNR: 25.76 -> 25.80 | Helped |
| Add MFDF to the configuration without OSFF modules (ablation) | OSFFNet · Urban100 ×4 | PSNR: 25.76 -> 25.86 | Helped |
| Add OISI and SFGC while omitting MFDF (ablation) | OSFFNet · Urban100 ×4 | PSNR: 25.76 -> 25.98 | Helped |
| Add OISI and MFDF while omitting SFGC (ablation) | OSFFNet · Urban100 ×4 | PSNR: 25.76 -> 26.12 | Helped |
| Add SFGC and MFDF while omitting OISI (ablation) | OSFFNet · Urban100 ×4 | PSNR: 25.76 -> 25.80 | No clear effect |
| Add OISI, SFGC, and MFDF (ablation) | OSFFNet · Urban100 ×4 | PSNR: 25.76 -> 26.12 | Helped |
Surprising findings
- Depth-extended IMDN outperforms the OSFF-based IMDN at x4 on several metrics, partly due to higher parameter count.
- OSFFNet does not achieve the best Set14 result at x2 scale.
Limitations
- Does not achieve best results on Set14 at x2 scale, indicating weaker local feature extraction.
- SFGC adds computational and memory overhead for lightweight networks.
- Gains from OSFF on IMDN are smaller than a parameter-matched depth-extended IMDN.
Experimental setup
Datasets: DIV2K Flickr2K
Benchmarks: Set5 Set14 BSD100 Urban100 Manga109
Metrics: PSNR SSIM #Params #Multi-Adds
Baselines: SRCNN VDSR EDSR CARN IMDN HNCT FMEN FDIWN BSRN RepRFN
Builds on: IMDN BSRN RepRFN SRCNN
Closest papers by meaning
- Residual Dense Network for Image Super-Resolution (CVPR 2018 · similarity 0.89)
- Omni Aggregation Networks for Lightweight Image Super-Resolution (CVPR 2023 · similarity 0.87)
- IAFMNet: Information-Aware Feature Modulation for Efficient Super-Resolution (CVPR 2026 · similarity 0.86)
- Lightweight Image Super-Resolution via Flexible Meta Pruning (ICML 2024 · similarity 0.84)
- SAVSR: Arbitrary-Scale Video Super-Resolution via a Learned Scale-Adaptive Network (AAAI 2024 · similarity 0.84)
- FIPER: Factorized Features for Robust Image Super-Resolution and Compression (NeurIPS 2025 · similarity 0.84)
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.