Research map / Label-efficient and robust learning / Self-supervised and contrastive
Self-Supervised Character-to-Character Distillation for Text Recognition
Tongkun Guan, Wei Shen, Xue Yang, Qi Feng, Zekun Jiang, Xiaokang 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 |
|---|---|---|---|
| use character-to-character distillation instead of sequence-level DiG | ViT-Base · STR avg, ARD fine-tuning | average accuracy: 94.92 -> 96.30 | Improved |
| use CCD instead of DiG with STD fine-tuning | ViT-Small · STR avg, STD fine-tuning | average accuracy: 93.23 -> 93.59 | Improved |
| add distillation strategy (ablation) | CCD · STR avg, ARD fine-tuning | average accuracy: 91.45 -> 95.57 | Improved |
| use geometric plus color augmentation on irregular view instead of color-only (R2R) (ablation) | CCD · STR avg, ARD fine-tuning | average accuracy: 95.08 -> 95.57 | Improved |
| train a self-supervised segmentation network instead of using K-means masks directly (ablation) | CCD · STR avg, ARD fine-tuning | average accuracy: 95.33 -> 95.57 | Improved |
| select pseudo-label cluster using border heuristic instead of arbitrary cluster (ablation) | text segmentation pseudo-labels · TextSeg | IoU (%): 40.1 -> 70.0 | Improved |
| train segmentation network on K-means pseudo-labels (ablation) | text segmentation · TextSeg | IoU (%): 70.0 -> 73.6 | Improved |
| use CCD pretraining instead of DiG for text segmentation | ViT-Small · TextSeg | IoU (%): 83.1 -> 84.8 | Improved |
| use CCD pretraining instead of DiG for super-resolution | ViT-Small · TextZoom | average SSIM: 0.7522 -> 0.7843 | Improved |
| evaluate frozen encoder features: CCD vs DiG | ViT-Small · 11 STR benchmarks, frozen encoder | average accuracy: 71.1 -> 69.9 | Hurt |
| use CCD with 10% label fraction instead of DiG | ViT-Small · 10% of ARD | average accuracy: 73.5 -> 78.2 | Improved |
| use CCD with 1% label fraction instead of DiG | ViT-Small · 1% of ARD | average accuracy: 62.9 -> 63.7 | Improved |
Closest papers by meaning
- CCDPlus: Towards Accurate Character to Character Distillation for Text Recognition (TPAMI 2025 · similarity 0.93)
- Self-Distillation Regularized Connectionist Temporal Classification Loss for Text Recognition: A Simple Yet Effective Approach (AAAI 2024 · similarity 0.86)
- OTE: Exploring Accurate Scene Text Recognition Using One Token (CVPR 2024 · similarity 0.86)
- Boosting Semi-Supervised Scene Text Recognition via Viewing and Summarizing (NeurIPS 2024 · similarity 0.85)
- Accurate Scene Text Recognition with Efficient Model Scaling and Cloze Self-Distillation (CVPR 2025 · similarity 0.85)
- Out of Length Text Recognition with Sub-String Matching (AAAI 2025 · 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.
