Research map / Image generation / Image restoration
Pyramid Attention Network for Image Restoration
Yiqun Mei, Yuchen Fan, Yulun Zhang, Jiahui Yu, Yuqian Zhou, Ding Liu, Yun Fu, Thomas S. Huang, Humphrey Shi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add single-scale non-local attention to the ResNet baseline (ablation) | ResNet 16 blocks · Urban100 denoising sigma=30 | PSNR: 30.86 -> 31.14 | Improved |
| replace non-local attention with pyramid attention (ablation) | ResNet 16 blocks · Urban100 denoising sigma=30 | PSNR: 31.14 -> 31.29; 0.43 dB over the baseline and 0.15 dB over classical non-local | Improved |
| use 3x3 block-wise matching instead of pixel-wise matching (ablation) | non-local attention · Urban100 denoising sigma=30 | PSNR: 31.14 dB to 31.21 dB | Improved |
| increase pyramid from 1 level to 5 levels (ablation) | pyramid attention · Urban100 denoising sigma=30 | PSNR: 31.21 -> 31.29 | Improved |
| insert pyramid attention at the middle position only (ablation) | ResNet · Urban100 denoising sigma=30 | PSNR: 30.86 -> 31.29 | Improved |
| insert pyramid attention at the pre position only (ablation) | ResNet · Urban100 denoising sigma=30 | PSNR: 30.86 -> 31.07 | Improved |
| insert pyramid attention at the post position only (ablation) | ResNet · Urban100 denoising sigma=30 | PSNR: 30.86 -> 31.18 | Improved |
| insert pyramid attention at pre, mid and post positions together (ablation) | ResNet · Urban100 denoising sigma=30 | PSNR: 31.29 -> 31.48 | Improved |
| add one pyramid attention block to a plain ResNet backbone instead of RNAN's multiple non-local blocks | PANet · Urban100 denoising | PSNR: more than 0.3 dB improvements over RNAN on all noisy levels | Improved |
| add pyramid attention to EDSR for super resolution | EDSR · Urban100 x4 SR | PSNR: 26.64 -> 27.01 | Improved |
| use a small 8-block PANet-S instead of RNAN (1LB+1NLB) | PANet-S · Urban100 denoising | PSNR: 28.36 -> 28.80 with 665K vs 1494K parameters | Improved |
| add pyramid attention to PANet for demosaicing versus RNAN | PANet · Urban100 demosaicing | PSNR: 39.75 -> 40.50 | Improved |
Closest papers by meaning
- Residual Non-local Attention Networks for Image Restoration (ICLR 2019 · similarity 0.88)
- KNN Local Attention for Image Restoration (CVPR 2022 · similarity 0.88)
- Accurate Image Restoration with Attention Retractable Transformer (ICLR 2023 · similarity 0.88)
- HAT: Hybrid Attention Transformer for Image Restoration (TPAMI 2025 · similarity 0.87)
- Efficient and Explicit Modelling of Image Hierarchies for Image Restoration (CVPR 2023 · similarity 0.87)
- Omni-Kernel Network for Image Restoration (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.
