Research map / Image generation / Image restoration
Image Restoration Using Very Deep Convolutional Encoder-Decoder Networks with Symmetric Skip Connections
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 symmetric skip connections (RED20 vs RED10, 10 vs 5 conv/deconv layers with skip) (ablation) | RED-Net deeper variant · 14 images denoising | PSNR: +0.12dB at sigma=10, +0.26dB at sigma=70 over RED10 | Helped |
| deepen network from 20 to 30 layers with skip connections (ablation) | RED-Net · 14 images denoising | PSNR: +0.07dB at sigma=10, +0.08dB at sigma=70 over RED20 | Helped |
| RED30 vs WNNM on 14-image denoising | RED-Net · 14 images denoising | PSNR: +0.32dB at sigma=10 up to +0.51dB at sigma=70 | Helped |
| RED30 vs WNNM on BSD200 denoising | RED-Net · BSD200 denoising | PSNR: +0.38dB, 0.47dB, 0.49dB, 0.42dB across sigma 10-70 | Helped |
| 30-layer vs CSCN on Set5 super-resolution | RED-Net · Set5 x2/x3/x4 | PSNR: +0.52dB, +0.56dB, +0.47dB | Helped |
| 30-layer vs CSCN on Set14 super-resolution | RED-Net · Set14 x2/x3/x4 | PSNR: +0.23dB, +0.06dB, +0.1dB | Helped |
| train a single 30-layer model across all noise levels instead of per-level models (ablation) | RED-Net · 14 images and BSD200 denoising | PSNR: slight degradation vs per-level models (e.g., 34.49 vs 34.81 at sigma=10) | Mixed |
| use element-wise skip connections (ours) instead of block-level residual connections of He et al. (ablation) | RED-Net · denoising training curves (sigma=70) | PSNR at last iteration: 25.30 (ours) vs 25.21 (He et al.) at last iteration of the curves | Helped |
| add skip connections to a 10-layer conv/deconv network (ablation) | RED-Net · training loss | training loss: lower loss with skip connections than without at 20 and 30 layers | Helped |
| use convolution/deconvolution with downsampling in conv layers (speed-up) at test time (ablation) | RED-Net · testing efficiency | PSNR | No clear effect |
| replace deconvolution decoder with plain convolution (fully convolutional, no skip) (ablation) | RED-Net decoder · denoising 5-layer vs 10-layer | PSNR | No clear effect |
Closest papers by meaning
- End-to-End Learning for Joint Image Demosaicing, Denoising and Super-Resolution (CVPR 2021 · similarity 0.90)
- Universal Denoising Networks : A Novel CNN Architecture for Image Denoising (CVPR 2018 · similarity 0.89)
- Dual Residual Networks Leveraging the Potential of Paired Operations for Image Restoration (CVPR 2019 · similarity 0.89)
- Deep Generalized Unfolding Networks for Image Restoration (CVPR 2022 · similarity 0.88)
- Deep Decoder: Concise Image Representations from Untrained Non-convolutional Networks (ICLR 2019 · similarity 0.88)
- Omni-Kernel Network for Image Restoration (AAAI 2024 · similarity 0.88)
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.