Research map / Detection and segmentation / Open-set and few-shot detection
General Detection-based Text Line 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 DTLR, a transformer-based character detection approach (DINO-DETR) for text line recognition that is synthetic-pretrained, masked, and fine-tuned with an adapted CTC loss on line-level labels, achieving SOTA on Chinese CASIA v2 and cipher datasets.
Problem
Text line recognition (OCR, HTR) across scripts is dominated by autoregressive or CTC-based recurrent methods; detection-based character localization has been discarded for handwriting because character-level annotations are scarce and characters are hard to separate.
Approach
Pretrain a DINO-DETR character detector on synthetic text lines with character-level boxes (with random erasing to encourage implicit context modeling), then fine-tune end-to-end on real line-level labels using a CTC variant that drops the blank-separation requirement, with the detector predicting all characters in parallel.
Key results
- On CASIA v2, DTLR reaches 96.83% AR and 97.34% CR, slightly above the prior best (96.77% AR / 96.77% CR).
- On the Copiale cipher, DTLR achieves 2.2% SER and 84.3% WA; on Borg it achieves 8.5% SER, beating the best prior baselines.
- On IAM, DTLR with N-gram reaches 5.46% CER, clearly behind DTrOCR (2.38%); on RIMES it reaches 2.53% CER with N-gram, competitive but behind the best method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add a character-level N-gram | DTLR · IAM | CER (%): 5.99 -> 5.46 | Helped |
| add a character-level N-gram | DTLR · READ | CER (%): 5.53 -> 5.35 | Helped |
| add a character-level N-gram | DTLR · RIMES | CER (%): 3.51 -> 2.53 | Helped |
| fine-tune the pre-trained detector on real data | DTLR · Google1000 English Volume 0002 | CER (%): 3.61% -> 2.04% | Helped |
| use English rather than the general pretraining model (ablation) | DTLR · IAM | CER (%): 39.72 -> 37.20 | Helped |
| add random erasing during pretraining (ablation) | DTLR English pretraining · IAM | CER (%): 37.20 -> 36.70 | Helped |
| freeze the network and train only the character classification layer instead of end-to-end fine-tuning (ablation) | DTLR · IAM | CER (%): 5.99 -> 20.48 | Hurt |
| freeze the network and train only the character classification layer instead of end-to-end fine-tuning (ablation) | DTLR · Borg | SER (%): 8.5 -> 49.87 | Hurt |
| add random erasing during fine-tuning (ablation) | DTLR · IAM | CER (%): 6.93 -> 5.99 | Helped |
| add random erasing during fine-tuning (ablation) | DTLR · Borg | SER (%): 13.1 -> 8.5 | Helped |
| use DTLR instead of the strongest listed baseline | DTLR · CASIA v2 | AR (%): 96.77 -> 96.83 | Helped |
| use DTLR instead of the strongest listed baseline | DTLR · CASIA v2 | CR (%): 96.77 -> 97.34 | Helped |
Surprising findings
- Random erasing during fine-tuning gives a large gain (Borg SER 13.1% to 8.5%, IAM CER 6.93% to 5.99%), suggesting it encourages an implicit language model.
- On CASIA v2, bounding boxes degenerate during long fine-tuning even as CER keeps improving, since no box-constraining loss is used.
Limitations
- Lags behind state-of-the-art on IAM and READ, attributed to lack of a strong language prior and small training sets.
- Bounding boxes can degenerate during long fine-tuning because no explicit box loss is used on real data.
- Ordering characters by minimum x-coordinate may fail for slanted scripts, though not observed in practice.
Experimental setup
Datasets: Google1000 IAM READ 2016 RIMES 2011 CASIA v1 CASIA v2 Borg cipher Copiale cipher Wikipedia sentences ICDAR 2024 cipher competition data
Benchmarks: Google1000 IAM READ 2016 RIMES 2011 CASIA v2 Borg cipher Copiale cipher ICDAR 2024 cipher competition
Metrics: Character Error Rate (CER) Accurate Rate (AR) Correct Rate (CR) Symbol Error Rate (SER) Word Accuracy (WA)
Baselines: TrOCR DTrOCR DAN PyLaia VAN FasterDAN Text-DIAE LSTM (VGG) Seq2Seq + Attention
Builds on: DINO-DETR CTC KenLM
Closest papers by meaning
- On the Generalization of Handwritten Text Recognition Models (CVPR 2025 · similarity 0.86)
- Out of Length Text Recognition with Sub-String Matching (AAAI 2025 · similarity 0.85)
- Reading between the Lines: Image-Based Order Detection in OCR for Chinese Historical Documents (AAAI 2024 · similarity 0.85)
- OCRBench v2: An Improved Benchmark for Evaluating Large Multimodal Models on Visual Text Localization and Reasoning (NeurIPS 2025 · similarity 0.84)
- OTE: Exploring Accurate Scene Text Recognition Using One Token (CVPR 2024 · similarity 0.84)
- MetaWriter: Personalized Handwritten Text Recognition Using Meta-Learned Prompt Tuning (CVPR 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.