Research map / Label-efficient and robust learning / Self-supervised and contrastive
Perceiving Stroke-Semantic Context: Hierarchical Contrastive Learning for Robust Scene Text Recognition
Hao Liu, Bin Wang, Zhimin Bao, Mobai Xue, Sheng Kang, Deqiang Jiang, Yinsong Liu, Bo Ren
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 hierarchical stroke- and semantic-context contrastive learning instead of cross-sample sequential contrast | Scene text recognition representations · IIIT5K, IC03, IC13 | word accuracy: 35.7 -> 43.4; 43.6 -> 50.6; 43.5 -> 51.2 | Improved |
| Pre-train PerSec on UTI-100M | PerSec-ViT · IIIT5K, IC03, IC13 | word accuracy: 38.4 -> 43.4; 46.2 -> 50.6; 46.7 -> 51.2 | Improved |
| Use a ViT-based rather than CNN-based encoder | PerSec · IIIT5K, IC03, IC13 | word accuracy: 39.2 -> 43.4; 47.8 -> 50.6; 48.2 -> 51.2 | Improved |
| Add the stroke context perceiver (ablation) | PerSec-CNN · IC13 | word accuracy: 86.3 -> 87.9; 86.0 -> 87.9 | Improved |
| Add the stroke context perceiver (ablation) | PerSec-ViT · IC13 | word accuracy: 87.5 -> 89.7; 88.0 -> 89.2 | Improved |
| Add the semantic context perceiver (ablation) | PerSec-CNN · IC13 | word accuracy: 86.0 -> 87.9; 86.7 -> 88.9 | Improved |
| Add the semantic context perceiver (ablation) | PerSec-ViT · IC13 | word accuracy: 87.2 -> 89.7; 87.9 -> 89.2 | Improved |
| Remove both stroke and semantic context perceivers (ablation) | PerSec-CNN · IC13 | word accuracy: 83.3 -> 87.9; 85.4 -> 88.9 | Hurt |
| Remove both stroke and semantic context perceivers (ablation) | PerSec-ViT · IC13 | word accuracy: 84.2 -> 89.7; 86.2 -> 89.2 | Hurt |
| Remove the quantizer and contrast input features directly with context-aggregated features (ablation) | PerSec · IC13 | convergence | Hurt |
| Use PerSec-pretrained encoders for existing recognizers | RobustScanner, SATRN, SAR, and NRTR · IC13, IIIT5K, SVT, IC15, SVTP, CT80 | word accuracy: 92.8 -> 95.1; 94.6 -> 95.2; 88.1 -> 91.2; 76.9 -> 78.1; 79.5 -> 81.2; 89.7 -> 91.3 | Improved |
Closest papers by meaning
- Context-Based Contrastive Learning for Scene Text Recognition (AAAI 2022 · similarity 0.88)
- Boosting Semi-Supervised Scene Text Recognition via Viewing and Summarizing (NeurIPS 2024 · similarity 0.86)
- Linguistics-aware Masked Image Modeling for Self-supervised Scene Text Recognition (CVPR 2025 · similarity 0.86)
- Text Gestalt: Stroke-Aware Scene Text Image Super-resolution (AAAI 2022 · similarity 0.86)
- OTE: Exploring Accurate Scene Text Recognition Using One Token (CVPR 2024 · similarity 0.85)
- ConText: Driving In-context Learning for Text Removal and Segmentation (ICML 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.
