Research map / Label-efficient and robust learning / Recognition and long-tail
An Exemplar-based Framework for Chinese Text Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DECTR, an exemplar-based Chinese text recognizer that discovers character exemplars from text-line images with a DETR-style decoder (weak supervision) and corrects predictions by retrieving labels of similar exemplars from a database.
Problem
Chinese text recognition in scene, web, document and handwritten images is hard due to a very large character lexicon and many morphologically similar characters, while standard sequence-based recognizers struggle to distinguish them.
Approach
Replace sequence modeling with character-level exemplar discovery (attention-based decoder learns character positions from text-line labels only) followed by exemplar retrieval against an exemplar bank, using a lightweight projection network trained with triplet loss and late fusion of distances to propagate labels.
Key results
- DECTR-L reaches 77.44% scene accuracy and 70.02% web accuracy, outperforming prior methods such as CCR-CLIP (71.31%) and SVTR-L (72.1%).
- Exemplar retrieval (DECTR-B) improves scene accuracy from 69.47% (EDN-B classification head) to 73.10%.
- Scaling the synthetic exemplar bank to 3.59M exemplars raises web-subset accuracy from 68.01% to 70.0%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add exemplar retrieval to EDN-B (ablation) | EDN-B · Scene, Web, Document, Handwriting | ACC: 69.47% -> 73.10%; 62.62% -> 65.62%; 98.31% -> 98.52%; 49.02% -> 52.87% | Helped |
| add direct ranking with threshold (0.7) to classifying exemplars (ablation) | EDN-B · Scene, Web, Document, Handwriting | ACC: 69.47% -> 72.32%; 61.90% -> 64.18%; 98.05% -> 98.47%; 48.98% -> 51.90% | Helped |
| add 10 retrieved exemplars per query during projection training (ablation) | DECTR-B projection network · Scene | ACC: Top 1 72.32% -> Top 10 73.10% | Helped |
| use RRT instead of the proposed feature combination strategy (ablation) | DECTR-B retrieval module · Scene | GFLOPs: 0.014 -> 0.225 | Hurt |
| increase encoder layers from 0 to 6 (ablation) | EDN-B · Scene | ACC: 59.39% -> 71.61% | Helped |
| increase encoder layers from 6 to 9 (ablation) | EDN-B · Scene | ACC: 71.61% -> 67.13% | Hurt |
| increase decoder layers from 1 to 4 (ablation) | EDN-B · Scene | ACC: 67.57% -> 69.65% | Helped |
| increase decoder layers from 4 to 6 (ablation) | EDN-B · Scene | ACC: 69.65% -> 67.63% | Hurt |
| add projection network to direct-ranking features (ablation) | DECTR-B retrieval module · Scene | ACC: 72.32% -> 73.10% | Helped |
| increase exemplar-bank size with synthesized exemplars (ablation) | DECTR-B · Web | ACC: 60.1% -> 70.0% | Helped |
Surprising findings
- Exemplar-based prediction via concatenated character labels is on par with state-of-the-art sequence-based methods, even those with more complex networks.
- Real-world web exemplars give higher accuracy than synthetic exemplars at similar bank size (68% vs 65%).
Limitations
- Gains on document images are minor because baseline accuracy is already high (~98%).
- Synthetic exemplar bank is less effective than real-world exemplars at equal size.
Experimental setup
Datasets: Chinese text recognition benchmark (Yu et al. 2021): scene, web, document, handwriting synthetic Chinese text-line images from oh-my ocr renderer
Benchmarks: Chinese text recognition benchmark (Yu et al. 2021) scene web document handwriting
Metrics: ACC NED GFLOPs #Params
Baselines: CRNN ASTER MORAN TransOCR ABINet SVTR-L MaskOCR-ViT-B CCR-CLIP
Builds on: DETR Deformable DETR image retrieval re-ranking ABINet
Closest papers by meaning
- TextScanner: Reading Characters in Order for Robust Scene Text Recognition (AAAI 2020 · similarity 0.85)
- Reading between the Lines: Image-Based Order Detection in OCR for Chinese Historical Documents (AAAI 2024 · similarity 0.84)
- OTE: Exploring Accurate Scene Text Recognition Using One Token (CVPR 2024 · similarity 0.83)
- Beyond Cropped Regions: New Benchmark and Corresponding Baseline for Chinese Scene Text Retrieval in Diverse Layouts (ICML 2025 · similarity 0.82)
- General Detection-based Text Line Recognition (NeurIPS 2024 · similarity 0.82)
- Out of Length Text Recognition with Sub-String Matching (AAAI 2025 · similarity 0.81)
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.