Research map / Image generation / Super-resolution
Mutual Affine Network for Spatially Variant Kernel Estimation in Blind Image Super-Resolution
Jingyun Liang, Guolei Sun, Kai Zhang, Luc Van Gool, Radu Timofte
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 group convolution with MAConv (2 splits) (ablation) | MANet · BSD100 x4 LR image | PSNR: 48.32 -> 49.39 | Improved |
| replace plain convolution with MAConv (2 splits) (ablation) | MANet · BSD100 x4 LR image | PSNR: 48.85 -> 49.39 | Improved |
| increase MAConv channel splits from 2 to 4 (ablation) | MANet · BSD100 x4 LR image | PSNR: 49.39 -> 49.77 | Improved |
| increase MAConv channel splits from 4 to 6 (ablation) | MANet · BSD100 x4 LR image | PSNR: 49.77 -> 49.80 | No clear effect |
| increase channels from [32,64,32] to [128,256,128] (ablation) | MANet with MAConv, 2 splits · BSD100 x4 LR image | PSNR: 45.87 -> 49.39 | Improved |
| use MAConv instead of group convolution at small width [32,64,32] (ablation) | MANet · BSD100 x4 LR image | PSNR: 45.14 -> 45.87 | Improved |
| increase MAConv layers per residual block from 2 to 4 (receptive field 22x22 to 38x38) (ablation) | MANet · toy spatially variant image | kernel accuracy | Hurt |
| train with kernel loss instead of LR image loss (ablation) | MANet · toy spatially variant image | kernel estimation | Improved |
| add Gaussian noise and JPEG compression to input (sigma 15, q 70 vs clean) | MANet · BSD100 x4 | LR image PSNR: 45.45 -> 40.59 | Hurt |
| estimate spatially variant kernels with MANet instead of one kernel per image with IKC | RRDB-SFT · BSD100 spatially variant type 2 (x4) | PSNR: 26.03 -> 27.21 | Improved |
| estimate per-patch kernels with MANet instead of IKC | RRDB-SFT · Urban100 spatially invariant (x4) | PSNR: 25.16 -> 25.66 | Improved |
| MANet kernels under heavy noise vs IKC | RRDB-SFT · BSD100 spatially variant type 1 (x4, noise 15) | PSNR: 24.64 -> 24.89 | Improved |
Closest papers by meaning
- Blind Super-Resolution With Iterative Kernel Correction (CVPR 2019 · similarity 0.88)
- Uncertainty Learning in Kernel Estimation for Multi-stage Blind Image Super-Resolution (ECCV 2022 · similarity 0.86)
- Learning the Non-Differentiable Optimization for Blind Super-Resolution (CVPR 2021 · similarity 0.86)
- Flow-Based Kernel Prior With Application to Blind Super-Resolution (CVPR 2021 · similarity 0.86)
- Spectrum-to-Kernel Translation for Accurate Blind Image Super-Resolution (NeurIPS 2021 · similarity 0.85)
- KOALAnet: Blind Super-Resolution Using Kernel-Oriented Adaptive Local Adjustment (CVPR 2021 · similarity 0.85)
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.
