Researchers have explored adapting the TrOCR model for handwritten text recognition (HTR) on medieval manuscripts, a task complicated by the model's pre-training on modern text. Through controlled experiments on a 13th-century Italian manuscript (I-CT 91 "Cortonese") and the READ-16 benchmark, they investigated the impact of contrast normalization, data augmentation, and layer freezing on accuracy. The study found that removing contrast normalization achieved a character error rate (CER) of 7.84%, comparable to a specialized baseline, and that specific layer freezing strategies could be transferred across datasets, though dataset-specific re-validation is advised. Grad-CAM and cross-attention maps were used to diagnose error patterns. AI
IMPACT This research offers insights into fine-tuning transformer models for specialized historical text recognition tasks.
RANK_REASON The cluster contains an academic paper detailing a systematic study and ablation of a model's performance on a specific task.
- Contrast Limited Adaptive Histogram Equalization
- Federico Simonetta
- Grad-CAM++
- Hugging Face
- I-CT 91 "Cortonese"
- READ-16
- TrOCR
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
