Research map / Label-efficient and robust learning / Test-time adaptation
MetaWriter: Personalized Handwritten Text Recognition Using Meta-Learned Prompt Tuning
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 MetaWriter, which personalizes handwritten text recognition at test time by tuning padding-based visual prompts with an unlabeled MAE (SSIM) reconstruction loss, with meta-learned prompt initialization.
Problem
Writer-specific personalization of handwritten text recognition at test time without labeled examples and without the cost of full-parameter gradient-based meta-learning adaptation.
Approach
Formulates writer adaptation as visual prompt tuning in which learnable padding vectors are updated in an inner loop using a self-supervised masked-image reconstruction (SSIM) loss on unlabeled support images, while the outer meta-learning loop optimizes the prompt initialization using a supervised CTC/cross-entropy text loss on labeled query images.
Key results
- On IAM line-level test set with 5-shot adaptation, MetaWriter achieves 3.36% CER and 10.32% WER, versus 4.32% CER and 16.24% WER for VAN.
- On RIMES line-level test set, MetaWriter achieves 2.19% CER and 6.63% WER, versus 2.63% CER and 6.78% WER for DAN.
- Only 0.08M parameters are tuned during personalization (about 1% of the model), compared with 7.6M for DAN and 1.7M for MetaHTR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MetaWriter for 5-shot unlabeled adaptation | HTR on IAM · IAM test set | WER: 16.24% -> 10.32% | Helped |
| use MetaWriter for 5-shot unlabeled adaptation | HTR on RIMES · RIMES test set | WER: 6.78% -> 6.63% | Helped |
| add the MAE reconstruction task (ablation) | MetaWriter · IAM | WER: 11.06% -> 10.21% | Helped |
| add the MAE reconstruction task (ablation) | MetaWriter · RIMES | WER: 7.24% -> 6.55% | Helped |
| add meta-learning for prompt initialization (ablation) | MetaWriter · IAM | WER: 10.75% -> 10.21% | Helped |
| add meta-learning for prompt initialization (ablation) | MetaWriter · RIMES | WER: 7.08% -> 6.55% | Helped |
| add prompts (ablation) | MetaWriter · IAM | WER: 10.22% -> 10.21% | Hurt |
| add prompts (ablation) | MetaWriter · RIMES | WER: 6.56% -> 6.55% | Helped |
| increase prompt coverage from 0 to 18 convolutional layers (ablation) | MetaWriter · IAM | WER: 10.63% -> 10.2% | Helped |
| increase prompt coverage from 0 to 18 convolutional layers (ablation) | MetaWriter · RIMES | WER: 6.89% -> 6.52% | Helped |
| increase support set from K=1 to K=5 (ablation) | MetaWriter · IAM | WER: 10.53% -> 10.32% | Helped |
| increase support set from K=1 to K=5 (ablation) | MetaWriter · RIMES | WER: 6.84% -> 6.63% | Helped |
Surprising findings
- Ablation shows removing the meta-learning component (w/o Meta) yields 3.63% CER on IAM, still better than the baseline, and removing prompts (w/o Prompt) gives the lowest error rates among ablations, at the cost of full parameter updates.
Limitations
- Evaluated only at line level; the authors note the method may not perform well at page level because padding prompts ignore text location and reading order.
- Direct comparison with MetaHTR is difficult due to differing settings (word-level WRA vs line-level CER/WER, labeled vs unlabeled adaptation).
Experimental setup
Datasets: IAM Handwriting Database RIMES2011 synthetic printed lines
Benchmarks: IAM Handwriting Database RIMES2011
Metrics: CER WER number of trainable parameters
Baselines: MetaHTR TrOCR SPAN DAN VAN Coquenet et al. Wang et al. Puigcerver et al.
Builds on: MetaHTR MAML Visual Prompt Tuning Masked Autoencoder DAN
Closest papers by meaning
- On the Generalization of Handwritten Text Recognition Models (CVPR 2025 · similarity 0.87)
- General Detection-based Text Line Recognition (NeurIPS 2024 · similarity 0.84)
- Muharaf: Manuscripts of Handwritten Arabic Dataset for Cursive Text Recognition (NeurIPS 2024 · similarity 0.84)
- AttnDreamBooth: Towards Text-Aligned Personalized Text-to-Image Generation (NeurIPS 2024 · similarity 0.83)
- CoRe: Context-Regularized Text Embedding Learning for Text-to-Image Personalization (AAAI 2025 · similarity 0.82)
- Zero-Shot Styled Text Image Generation, but Make It Autoregressive (CVPR 2025 · similarity 0.82)
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.