Research map / Vision-language and multimodal / CLIP and zero-shot
Chinese Text Recognition with A Pre-Trained CLIP-Like Model Through Image-IDS Aligning
Haiyang Yu, Xiaocong Wang, Bin Li, Xiangyang Xue
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 matching head with canonical representations instead of fully connected prediction head (ablation) | TransOCR-style CTR model · CTR benchmark Handwriting | LACC: 53.00% -> 58.54% | Improved |
| add matching head (ablation) | TransOCR-style CTR model · CTR benchmark Scene | LACC: 71.33% -> 70.17% | Hurt |
| add regularization term to L_ctr (ablation) | CTR model with matching head · CTR benchmark Handwriting | LACC: 58.54% -> 60.30% | Improved |
| add image-image contrastive loss L_I (lambda=1 vs 0) (ablation) | CCR-CLIP · CTW | CACC: 83.29% -> 85.78% | Improved |
| use radical-level instead of stroke-level text representation (ablation) | CCR-CLIP text encoder · ICDAR2013 | CACC: 92.74% -> 97.18% | Improved |
| use radical-level instead of character-level representation (ablation) | CCR-CLIP text encoder · CTW | CACC: 82.73% -> 85.78% | Improved |
| choose regularization weight beta=0.001 versus 0 or larger (ablation) | CTR model · CTR benchmark Handwriting | LACC: 59.54% -> 60.30% | Improved |
| increase beta from 0.001 to 1 (ablation) | CTR model · CTR benchmark Handwriting | LACC: 60.30% -> 58.76% | Hurt |
| image-IDS matching framework for zero-shot characters in text lines | CTR model vs TransOCR · CTR benchmark zero-shot characters | character accuracy: 51.4% average vs TransOCR cannot recognize them | Improved |
| image-IDS matching for few-shot characters | CTR model vs TransOCR · CTR benchmark few-shot characters | character accuracy: +10.6% average | Improved |
| image-IDS pre-training vs CUE under no printed images | CCR-CLIP · HWDB character zero-shot | CACC: +28.37% average | Improved |
| apply method to scene text with vertical text | Ours · CTR benchmark Scene | LACC: 71.33 -> 71.31 | No clear effect |
Closest papers by meaning
- MSA2: Multi-task Framework with Structure-aware and Style-adaptive Character Representation for Open-set Chinese Text Recognition (ICCV 2025 · similarity 0.88)
- TextScanner: Reading Characters in Order for Robust Scene Text Recognition (AAAI 2020 · similarity 0.87)
- An Exemplar-based Framework for Chinese Text Recognition (AAAI 2025 · similarity 0.87)
- General Detection-based Text Line Recognition (NeurIPS 2024 · similarity 0.85)
- Reading between the Lines: Image-Based Order Detection in OCR for Chinese Historical Documents (AAAI 2024 · similarity 0.85)
- OTE: Exploring Accurate Scene Text Recognition Using One Token (CVPR 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.
