Research map / Image generation / Super-resolution
Accelerating the Super-Resolution Convolutional Neural Network
Chao Dong, Chen Change Loy, Xiaoou Tang
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 |
|---|---|---|---|
| replace last conv layer with a deconvolution layer on the LR input (ablation) | SRCNN-Ex · Set5 x3 | PSNR (dB): 32.83 -> 32.95 | Improved |
| replace the 5x5 mapping layer with shrinking, 4 narrow 3x3 mapping layers and expanding (ablation) | SRCNN-Ex with deconvolution · Set5 x3 | PSNR (dB): 32.95 -> 33.01 | Improved |
| use smaller first-layer filters and fewer filters (Conv(5,56,1)) (ablation) | FSRCNN · Set5 x3 | PSNR (dB): 33.01 -> 33.06 | Improved |
| deconvolution on LR input instead of conv on upsampled input (ablation) | SRCNN-Ex · Set5 x3 | speedup: 1x -> 8.7x | Improved |
| add the expanding layer after the mapping layers (ablation) | FSRCNN · Set5 | PSNR (dB): decreases up to 0.3 dB without it | Improved |
| learn diverse deconvolution kernels per feature map instead of forcing identical kernels (ablation) | FSRCNN · Set5 | PSNR (dB): drops at least 0.9 dB when identical | Improved |
| increase mapping depth from m=2 to m=4 (d=48, s=12) (ablation) | FSRCNN · Set5 | PSNR (dB): 32.87 -> 33.08 | Improved |
| increase d and s from (48,12) to (56,16) at m=4 (ablation) | FSRCNN · Set5 | PSNR (dB): 33.08 -> 33.17 | No clear effect |
| smallest network FSRCNN (48,12,2) vs SRCNN-Ex | FSRCNN · Set5 | PSNR (dB): 32.75 -> 32.87 | Improved |
| transfer conv filters from x3 and fine-tune only the deconvolution for x2 (ablation) | FSRCNN (56,12,4) · x2 training | convergence / PSNR | No clear effect |
| use single FSRCNN network for x4 instead of SCN's two x2 models | FSRCNN (56,12,4) · Set5 x4 | PSNR (dB): 30.82 -> 30.55 | Hurt |
| FSRCNN-s (32,5,1) real-time model vs SRCNN | FSRCNN-s · Set5 x3 | PSNR (dB): 32.45 -> 32.55 | Improved |
Closest papers by meaning
- Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network (CVPR 2016 · similarity 0.94)
- Image Super-Resolution Using Deep Convolutional Networks (TPAMI 2016 · similarity 0.93)
- "Restore Globally, Refine Locally: A Mask-Guided Scheme to Accelerate Super-Resolution Networks" (ECCV 2022 · similarity 0.92)
- Learning with Privileged Information for Efficient Image Super-Resolution (ECCV 2020 · similarity 0.91)
- Accurate Image Super-Resolution Using Very Deep Convolutional Networks (CVPR 2016 · similarity 0.91)
- Deep Learning for Image Super-Resolution: A Survey (TPAMI 2020 · similarity 0.91)
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.
