Research map / Image generation / Text-to-image diffusion
Out of Length Text Recognition with Sub-String Matching
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 SMTR, a sub-string matching scene text recognizer trained only on short text that recognizes arbitrarily long text by iteratively predicting next/previous characters, plus a Long Text Benchmark (LTB) for out-of-length recognition.
Problem
Scene text recognizers trained on short word-level data fail to recognize long, multi-word text lines (out-of-length text), and existing benchmarks lack long-text evaluation.
Approach
Recognize text by matching a sub-string within the image via two cross-attention modules that encode the sub-string into next/previous queries and locate its neighboring characters, iterating to cover arbitrary length; add regularization training with randomly replaced characters to separate similar sub-strings and an inference augmentation that slices long images to handle repeated sub-strings.
Key results
- On LTB (trained only on Union14M-L short text), SMTR w/ IA reaches 50.99% W-Avg and 41.57% A-Avg, versus 42.56% W-Avg for LISTER and 0% for PARSeq.
- SMTR attains 85.00% average accuracy on Union14M-Benchmark and 95.90% on common benchmarks, comparable to PARSeq (84.26% and 96.40%).
- On the Chinese CTR Scene subset SMTR w/o Aug gets 79.8% vs LISTER 73.0%, and the overall CTR average reaches 82.68% with augmentation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add regularization training with 2 regularized sub-strings (ablation) | SMTR · LTB | accuracy: 32.30 -> 47.00 | Helped |
| increase regularized sub-strings from 1 to 2 (ablation) | SMTR · LTB | accuracy: 44.37 -> 47.00 | Helped |
| set sub-string length to 4 instead of 5 (ablation) | SMTR · LTB | accuracy: 47.00 -> 45.24 | Hurt |
| set sub-string length to 6 instead of 5 (ablation) | SMTR · LTB | accuracy: 47.00 -> 45.06 | Hurt |
| set sub-string length to 7 instead of 5 (ablation) | SMTR · LTB | accuracy: 47.00 -> 40.55 | Hurt |
| add inference augmentation (ablation) | SMTR · LTB W-Avg | accuracy: 47.00 -> 50.99 | Helped |
| add inference augmentation (ablation) | SMTR · LTB A-Avg | accuracy: 34.42 -> 41.57 | Helped |
| add inference augmentation (ablation) | SMTR · LTB [26, 35] | accuracy: +2.42% | Helped |
| add inference augmentation (ablation) | SMTR · LTB [36, 55] | accuracy: +6.63% | Helped |
| add inference augmentation (ablation) | SMTR · LTB ≥ 56 | accuracy: +12.4% | Helped |
| remove the matcher residual and MLP (ablation) | SMTR · LTB | accuracy: 43.33 -> 47.00 | Helped |
| set matcher heads to 2 instead of 12, with no residual or MLP (ablation) | SMTR · LTB | accuracy: 44.74 -> 47.00 | Helped |
Surprising findings
- PARSeq scores 0% on LTB despite strong short-text accuracy, showing absolute positional decoders fail beyond training length.
- Removing residual and MLP from the sub-string matcher improved long-text recognition, while hurting short-text performance slightly.
Limitations
- Iterative next/previous decoding makes inference slower than parallel decoders.
- Repeated sub-strings within the same text image still cause recognition errors, with 0.4% of LTB samples affected after IA.
Experimental setup
Datasets: Union14M-L Union14M-Benchmark IC13 SVT IIIT5K IC15 SVTP CUTE80 CTR
Benchmarks: LTB IC13 SVT IIIT IC15 SVTP CUTE Union14M-Benchmark CTR
Metrics: word accuracy
Baselines: PARSeq LISTER FocalSVTR AR-STR NRTR CDistNet CAM CCR-CLIP
Builds on: SVTR FocalSVTR LISTER PARSeq Union14M
Closest papers by meaning
- OTE: Exploring Accurate Scene Text Recognition Using One Token (CVPR 2024 · similarity 0.87)
- General Detection-based Text Line Recognition (NeurIPS 2024 · similarity 0.85)
- OCRBench v2: An Improved Benchmark for Evaluating Large Multimodal Models on Visual Text Localization and Reasoning (NeurIPS 2025 · similarity 0.85)
- On the Generalization of Handwritten Text Recognition Models (CVPR 2025 · similarity 0.84)
- An Empirical Study of Scaling Law for Scene Text Recognition (CVPR 2024 · similarity 0.84)
- LoTLIP: Improving Language-Image Pre-training for Long Text Understanding (NeurIPS 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.