Research map / Image generation / Super-resolution
Learning to Super-Resolve Blurry Face and Text Images
Xiangyu Xu, Deqing Sun, Jinshan Pan, Yujin Zhang, Hanspeter Pfister, Ming-Hsuan Yang
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 feature matching loss (G2D3) to basic GAN loss (ablation) | SCGAN · Face dataset | PSNR (dB): 22.82 -> 24.57 | Improved |
| use feature matching loss (G2D3) instead of pixel-wise loss (ablation) | SCGAN · Face dataset | PSNR (dB): 25.12 -> 24.57 (DCR 1.1331 -> 1.0687) | Mixed |
| use discriminator features from layer 3 for D update (G2D3) instead of layer 2 (G2D2) (ablation) | SCGAN · Face dataset | PSNR (dB): 22.95 -> 24.57 | Improved |
| use layer 3 features for generator update (G3D3) instead of layer 2 (G2D3) (ablation) | SCGAN · Face dataset | PSNR (dB): 24.57 -> 23.17 | Hurt |
| split batches into real and generated when training discriminator (ablation) | SCGAN · text and face images | visual quality | Improved |
| use bicubic interpolation instead of deconvolution for upsampling (ablation) | SCGAN generator · synthetic face dataset | PSNR (dB): 24.57 -> 23.45 | Hurt |
| train one multi-class model (MCGAN) instead of per-class SCGAN | GAN restoration · Text dataset | PSNR (dB): 20.65 -> 20.12 | Hurt |
| train one multi-class model (MCGAN) instead of per-class SCGAN | GAN restoration · Face dataset | PSNR (dB): 24.57 -> 23.95 | Hurt |
| use SCGAN instead of best sequential deblur+SR combination | blurry low-res restoration · Text dataset | PSNR (dB): 16.45 -> 20.65 | Improved |
| use SCGAN instead of best sequential deblur+SR combination | blurry low-res restoration · Face dataset | PSNR (dB): 22.17 -> 24.57 | Improved |
| use SCGAN instead of MSE-trained fine-tuned CNN ([18]+[15]) | blurry low-res restoration · Face dataset | PSNR (dB): 22.60 -> 24.57 | Improved |
| use SCGAN instead of bicubic in subjective study | blurry low-res restoration · Text images user study | mean opinion score: 1.10 -> 3.68 | Improved |
Closest papers by meaning
- To learn image super-resolution, use a GAN to learn how to do image degradation first (ECCV 2018 · similarity 0.93)
- Finding Tiny Faces in the Wild With Generative Adversarial Network (CVPR 2018 · similarity 0.90)
- Image-Adaptive GAN Based Reconstruction (AAAI 2020 · similarity 0.89)
- Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network (CVPR 2017 · similarity 0.89)
- GAN Prior Embedded Network for Blind Face Restoration in the Wild (CVPR 2021 · similarity 0.89)
- GLEAN: Generative Latent Bank for Large-Factor Image Super-Resolution (CVPR 2021 · similarity 0.89)
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.
