Research map / Image generation / Image restoration
The Devil is in the Upsampling: Architectural Decisions Made Simpler for Denoising with Deep Image Prior
Yilin Liu, Jiang Li, Yunkui Pang, Dong Nie, Pew-Thian Yap
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 |
|---|---|---|---|
| use a 2-level hourglass with per-image estimated width instead of the 5-level DIP at high noise | DIP hourglass · Set12 sigma=50 Gaussian | PSNR: 22.15 -> 24.59 | Improved |
| use the estimated under-parameterized network at mild noise | DIP hourglass · CBSD68 sigma=25 Gaussian | PSNR: 28.93 -> 28.57 | No clear effect |
| estimated architecture on real noise versus DIP | DIP hourglass · PolyU | PSNR: 38.15 -> 38.05 | No clear effect |
| remove three upsampling layers and scale width per image | ConvDecoder · CBSD68 sigma=25 | PSNR: 25.19 -> 28.29 | Improved |
| same ConvDecoder modification at higher noise | ConvDecoder · CBSD68 sigma=50 | PSNR: 24.36 -> 24.12 | No clear effect |
| remove upsampling from the decoder (ablation) | MLP-Decoder / ConvDecoder · Image denoising (Set9) | denoising PSNR | Hurt |
| replace bilinear with transposed convolution upsampling (ablation) | decoder · Image denoising (Set9) | PSNR over iterations | Hurt |
| use NN upsampler instead of bilinear on a coarse image (last iteration) (ablation) | ConvDecoder · Set9 coarse image | PSNR: 31.1 -> 27.8 | Hurt |
| use NN upsampler instead of bilinear on a fine image (ablation) | ConvDecoder · Set9 fine image (peak) | PSNR: 22.9 -> 24.5 | Improved |
| use an LPF with about 100dB attenuation in the upsampler (ablation) | ConvDecoder · Set9 coarse image (peak) | PSNR: 31.3 -> 27.5 | Hurt |
| increase network depth (ablation) | hourglass DIP · fine-grained images | SSIM | Hurt |
| add skip connections to deep network (ablation) | hourglass DIP · fine-grained images | SSIM | Improved |
Closest papers by meaning
- Rethinking Deep Image Prior for Denoising (ICCV 2021 · similarity 0.91)
- Universal Denoising Networks : A Novel CNN Architecture for Image Denoising (CVPR 2018 · similarity 0.90)
- On Measuring and Controlling the Spectral Bias of the Deep Image Prior (IJCV 2022 · similarity 0.89)
- Towards Architecture-Agnostic Untrained Networks Priors for Image Reconstruction with Frequency Regularization (ECCV 2024 · similarity 0.89)
- ISNAS-DIP: Image-Specific Neural Architecture Search for Deep Image Prior (CVPR 2022 · similarity 0.89)
- Rethinking Reconstruction and Denoising in the Dark: New Perspective, General Architecture and Beyond (CVPR 2025 · 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.
