Research map / Image generation / Super-resolution
Text Gestalt: Stroke-Aware Scene Text Image Super-resolution
Jingye Chen, Haiyang Yu, Jianqi Ma, Bin Li, Xiangyang Xue
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 stroke-level focus instead of no focus (ablation) | TSRN · TextZoom test | average recognition accuracy (CRNN): 41.4% -> 48.9% | Improved |
| Replace character-level focus with stroke-level focus (ablation) | TSRN · TextZoom test | average recognition accuracy (CRNN): 42.9% -> 48.9% | Improved |
| Add stroke-level focus instead of no focus (ablation) | TSRN · TextZoom test | average recognition accuracy (MORAN): 55.4% -> 59.4% | Improved |
| Add stroke-level focus instead of no focus (ablation) | TSRN · TextZoom test | average recognition accuracy (ASTER): 58.3% -> 61.3% | Improved |
| Add stroke-level focus instead of character-level focus (ablation) | TSRN · IC15-352 | recognition accuracy (CRNN): 61.1% -> 64.2% | Improved |
| Add stroke-level focus instead of no focus (ablation) | TSRN · Degraded-IC13 | recognition accuracy: 83.4% -> 86.1% | Improved |
| Add stroke-level focus instead of character-level focus (ablation) | TSRN · PE92 | recognition accuracy: 90.32% -> 92.37% | Improved |
| Use stroke-level attention maps only for correct predictions rather than all predictions (ablation) | TSRN · TextZoom test | average recognition accuracy (CRNN): 48.9% v.s. 48.5% | No clear effect |
| Use attention maps from wrong predictions (ablation) | TSRN · TextZoom test | average recognition accuracy (CRNN): 48.9% -> 39.8% | Hurt |
| Set λ_SFM to 50 rather than 0 (ablation) | TSRN · TextZoom test | average recognition accuracy (CRNN): 41.4% -> 48.9% | Improved |
| Set λ_SFM to 100 rather than 50 (ablation) | TSRN · TextZoom test | average recognition accuracy (CRNN): 48.9% -> 48.4% | Hurt |
| Use L1 rather than L2 for the SFM loss while keeping L2 PSM (ablation) | TSRN + SFM · TextZoom test | average recognition accuracy (CRNN): 47.0% -> 48.9% | Improved |
Closest papers by meaning
- Enhanced Generative Structure Prior for Chinese Text Image Super-Resolution (TPAMI 2025 · similarity 0.88)
- PRISM: Prior Rectification and Uncertainty-Aware Structure Modeling for Diffusion-Based Text Image Super-Resolution (NeurIPS 2026 · similarity 0.86)
- DreamText: High Fidelity Scene Text Synthesis (CVPR 2025 · similarity 0.85)
- Text-Aware Real-World Image Super-Resolution via Diffusion Model with Joint Segmentation Decoders (NeurIPS 2025 · similarity 0.85)
- OTE: Exploring Accurate Scene Text Recognition Using One Token (CVPR 2024 · similarity 0.84)
- Boosting Semi-Supervised Scene Text Recognition via Viewing and Summarizing (NeurIPS 2024 · similarity 0.84)
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.
