Research map / Label-efficient and robust learning / Knowledge distillation
Text Is Text, No Matter What: Unifying Text Recognition Using Knowledge Distillation
Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Yi-Zhe Song
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 logits distillation loss (ablation) | unified student · IC15 | word recognition accuracy (%): 70.4 -> 75.3 | Improved |
| add logits distillation loss (ablation) | unified student · IAM | word recognition accuracy (%): 81.8 -> 84.9 | Improved |
| add attention distillation loss (ablation) | unified student · IC15 | word recognition accuracy (%): 75.3 -> 75.7 | Improved |
| add hint distillation loss (ablation) | unified student · IC15 | word recognition accuracy (%): 75.7 -> 76.4 | Improved |
| add affinity distillation loss (ablation) | unified student · IC15 | word recognition accuracy (%): 76.4 -> 76.9 | Improved |
| add affinity distillation loss (ablation) | unified student · IAM | word recognition accuracy (%): 85.9 -> 86.4 | Improved |
| add conditional distillation (ablation) | unified student · IC15 | word recognition accuracy (%): reduced by 2.5% without it | Improved |
| add conditional distillation (ablation) | unified student · IAM | word recognition accuracy (%): reduced by 0.4% without it | Improved |
| vary omega from 1.01 to 1.10 (ablation) | unified student · IC15 | word recognition accuracy (%): 76.8 / 76.9 / 76.9 / 76.8 / 76.8 | No clear effect |
| put hint loss on context vector instead of CNN features (ablation) | unified student · IC15 | word recognition accuracy (%): 3.8% -> 5.1% improvement | Improved |
| put hint loss on context vector instead of decoder hidden state (ablation) | unified student · IC15 | word recognition accuracy (%): 4.6% -> 5.1% improvement | Improved |
| add image rectification module to KD pipeline (ablation) | unified student · IC15 | word recognition accuracy (%): 76.9 -> 75.9 | Hurt |
Closest papers by meaning
- Handwritten Text Recognition: A Survey (TPAMI 2025 · similarity 0.90)
- OTE: Exploring Accurate Scene Text Recognition Using One Token (CVPR 2024 · similarity 0.88)
- OCRBench v2: An Improved Benchmark for Evaluating Large Multimodal Models on Visual Text Localization and Reasoning (NeurIPS 2025 · similarity 0.87)
- Out of Length Text Recognition with Sub-String Matching (AAAI 2025 · similarity 0.87)
- Boosting Semi-Supervised Scene Text Recognition via Viewing and Summarizing (NeurIPS 2024 · similarity 0.87)
- On the Generalization of Handwritten Text Recognition Models (CVPR 2025 · similarity 0.87)
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.
