Source-linked AI summary
Leveraging Morphology for Historical Script Metrological Analysis
Malamatenia Vlachou Efstathiou, Raphaël Baena, Dominique Stutzmann, Mathieu Aubry
TL;DR
Paleographic analysis lacks scalable, reproducible access to interpretable visual measurements. The paper combines detection-based recognition with prototype-based reconstruction to learn character shapes and geometry, and shows that the resulting measurements support hand differentiation and subtle variation analysis in a four-hand codex.
Problem
Paleographic measurement remains limited by subjective expert comparison and by prior methods that do not provide standardized, trustworthy, scalable handwriting features.
Method
The paper extends detection-based text recognition with prototype-based line reconstruction, using learned character prototypes and bounding-box alignment to measure character shapes and spatial relationships.
Results
The framework produces more stable training, higher-quality prototypes, and scalable meaningful measurements that identify hands and reveal subtle paleographic variation in a 160-folio case study.
Takeaways & Limitations
Prototype alignment provides a consistent bridge between morphological and metrological analysis for fine-grained paleographic study.
Takeaways & Limitations
The prior Learnable Handwriter baseline lacks aspect-ratio changes and learns background and context, motivating the proposed architecture; refined canvas boxes are optional and mainly serve human perceptual fit.
Abstract
from arXiv · showhide
Advances in handwritten text recognition have enabled large-scale transcription of historical documents, but still provide limited access to interpretable visual measurements for paleography, the study of historical scripts. In this paper, our main insight is that morphological script analysis, in particular the capacity to learn character prototypes from line-level transcriptions, enables the definition of scalable, meaningful, and stable paleographic measurements. More precisely, we leverage a transformer-based detection architecture together with a prototype-based line reconstruction module to learn prototypical characters and their occurrence, deformation, and positioning. Our contributions are twofold. First, we introduce a deep architecture and learning methodology that enables efficient character modeling with only line-level transcription supervision, significantly improving over the Learnable Typewriter baseline and enabling accurate character bounding box prediction, unlocking its potential for paleographic measurements. Second, we introduce and demonstrate the paleographical relevance of automatic measurements enabled by our architecture for characters, bi-grams, and spaces between graphical units. For this demonstration, we extend the annotations of the codex Paris, BnF, fr. 2813, commissioned in the late fourteenth century by Charles V and copied by four hands, to 160 pages. We visualize our measurements over these pages, showing how they enable us not only to differentiate graphical profiles, but also to discover and analyze subtle variations. This case study outlines the scalability of our approach and its frugality in terms of required training data, since a single column of text is sufficient to compute our measurements on each of the 160 pages. Data and code are publicly available at: https://malamatenia.github.io/morphology4metrology-analysis.
1 Introduction
The paper addresses subjective, difficult-to-scale paleographic measurement by combining morphological character modeling with metrological analysis. It introduces a detection-and-reconstruction framework and demonstrates prototype-alignment measurements on a four-hand medieval codex.
- Paleographic analysis remains subjective and difficult to communicate, reproduce, and scale because it depends heavily on experts comparing remembered forms.
- The authors define character boundaries through optimal alignment between learned prototypes and specific character instances, linking morphology to metrology.
- The approach improves prototype quality and training stability over Learnable Handwriter while enabling paleographic measurements.
- The architecture extends detection-based recognition with prototype reconstruction, bounding-box deformation, accent and abbreviation modeling, and memory-efficient composition.
- On the Grandes Chroniques de France, visualizations identify four hands and reveal subtle inter- and intra-hand variations.
- The contributions establish scalable prototype-alignment measurements and demonstrate their potential for fine-grained analysis beyond hand recognition.
2 Related Work
Prior work spans sequence recognition, detection-based recognition, prototype reconstruction, and manual or semi-automatic paleographic measurement. This paper combines these strands to obtain interpretable prototypes, geometric measurements, and scalable script analysis.
- Detection-based recognition predicts character locations and geometry, providing information needed for fine-grained paleographic analysis beyond transcription-only models.
- Learnable Typewriter and Learnable Handwriter learn character prototypes from line images and transcriptions, but lack deformation-based measurements of distances, dimensions, and aspect ratios.
- The proposed framework extends DTLR with prototype-based reconstruction to obtain interpretable prototypes and metrological measurements in one system.
- Traditional and digital paleographic measurements have faced subjectivity, methodological inconsistency, and limited reproducibility.
- Existing approaches include prototype extraction, forced alignment, and statistical analysis of character proportions and spacing.
- Inter-character and inter-word spacing remains comparatively neglected, despite its relevance to pauses, phrasing, rhythm, and script density.
3 Methodology
The method reconstructs text lines from learned character prototypes using detection-based predictions for character identity, placement, deformation, accents, abbreviations, and color.
- The approach combines DTLR detection with prototype-based line reconstruction, using predicted bounding boxes to deform and place learned character prototypes.
- Each decoder token predicts character boxes, classes, accent and abbreviation information, and RGB color vectors.
- Learnable grayscale prototypes and a CNN-predicted low-resolution background support reconstruction alongside classification and box prediction.
- Bounding boxes enable aspect-ratio changes, prevent repeated prototype use, and reduce reconstruction cost by compositing nearby prototypes efficiently.
- Training assumes transcribed text lines are divided into meaningful subsets that serve as units of analysis for comparison.
Training Strategy.
Training combines detection and reconstruction objectives in a three-stage procedure that progresses from synthetic bounding-box learning to dataset-wide modeling and subset-specific refinement.
- The second stage trains a base model on all dataset lines using modified DTLR CTC loss together with L1 line-image reconstruction loss.
- The first stage pretrains the architecture on synthetic data using DINO-DETR detection losses to learn accurate character bounding boxes.
- The procedure follows three stages combining DTLR and Learnable Handwriter training strategies.
- The implementation uses a ResNet50-based architecture, a CNN background predictor, 48×48 grayscale prototypes, and a 78-character set.
Implementation and training details.
The method defines character boxes through prototype alignment rather than human annotations, while excluding unstable decorative extensions and refining canvas boundaries for measurement consistency.
- Prototype-aligned bounding boxes: Bounding boxes are selected to maximize alignment between each character instance and the dataset-wide learned prototype.The prediction head and backbone remain frozen during final fine-tuning, making boxes consistent across subsets.
- Prototype-aligned bounding boxes: Unlike the Learnable Typewriter, the alignment procedure permits character aspect-ratio changes.
- Boundary stabilization: The boxes exclude overextended attack and exit strokes, reducing variability from decorative or contextual elongations.This focuses measurements on stable character structure rather than outermost stroke contours.
- Boundary stabilization: A tighter prototype-canvas box is transported to character instances and used for measurements instead of the predicted boxes.It is on average 2.8 pixels smaller than the 48 × 48 canvas.
Refined boxes for improved consistency with human perception.
The refined bounding box is an optional perceptual adjustment that tightens prototype borders without changing the qualitative results.
- Refined boxes: The tighter bounding-box step is not necessary and does not change the qualitative results.It is added solely to better match human intuition about character borders.
Discarding errors.
Because recognition errors remain possible after fine-tuning, the method aligns predicted and ground-truth transcriptions to identify errors before measurement computation.
- Error filtering: The method discards recognition errors before computing measurements by comparing predicted and ground-truth transcriptions.Dynamic programming minimizes Levenshtein distance, similarly to Character Error Rate computation.
- Measurement inputs: The resulting character bounding boxes support measurements for characters, bigrams, and word separation.
Measures.
The analysis measures character proportions and horizontal spacing across units of analysis, normalizing distances for comparability and visualizing results with linear and crossed graphs.
- Measured properties: The measures include character width, character aspect ratio, bigram distance, bigram-box aspect ratio, and consecutive-word distance.Each measure is summarized by its mean and coefficient of variation across the unit of analysis.
- Measured properties: The coefficient of variation is computed as CV = σ/µ, with σ denoting standard deviation.
- Normalization: Distances are normalized with a unit based on half the average width of the lowercase m.This reference is intended to remain comparable across manuscripts with different text sizes and image resolutions.
- Dataset: The dataset summary contains 160 units of analysis from Paris, BnF, fr. 2813.
- Visualization: The measures are visualized using linear graphs and crossed graphs.
Visualisation types.
The paper uses linear graphs to track one measure across ordered analysis units and crossed graphs to examine relationships between pairs of measures. Together, they support trend detection, correlation analysis, script differentiation, and robustness assessment.
- Linear graphs display the evolution of a single measure across ordered units of analysis, such as pages or manuscripts arranged by date.
- Crossed graphs plot two measures against each other to analyze correlations or differentiate scripts using complementary measures.
- Marker intensity in both graph types indicates the number of occurrences contributing to each metric, revealing less robust data points.
4 Experiments
Experiments on the Grandes Chroniques de France evaluate the model against the Learnable Handwriter and apply its measurements to characterize graphic profiles and subtle variation. The results show more efficient, stable prototype learning and informative analyses of proportionality and spacing.
- 4.1 Case study and dataset: The case study uses the late fourteenth-century codex Paris, BnF, fr. 2813, copied by four hands and containing four graphic profiles.
- 4.2 Morphological analysis: 12.7 GB versus 31.3 GB: with batch size 128, the proposed architecture uses less memory than the Learnable Handwriter baseline.
- 4.2 Morphological analysis: 1 day and 20 hours versus 5 days and 2 hours: the proposed method reaches convergence faster, although the baseline achieves a slightly lower final CER of 1.2% versus 1.4%.The baseline’s converged prototypes shrink and accumulate contextual noise, so morphological analysis requires an earlier checkpoint with CER 4.3%.
- 4.2 Morphological analysis: The proposed prototypes are tightly fitted and centered, while baseline prototypes include neighboring strokes and inconsistent centering.Tight-box filtering removes an average of 2.8 pixels from the proposed boxes versus 18.4 pixels for the baseline.
- 4.3 Novel metrological analysis: Linear and crossed visualizations distinguish graphic profiles and reveal proportionality, executional consistency, and horizontal-compression patterns.Letter aspect ratios separate the four profiles; GP1 shows the lowest variability for ‹t›, while GP2 and GP3 show greater variability.
- 4.3 Novel metrological analysis: Bigram distances are consistently negative whereas word distances are positive, and word spacing decreases early in GP1 before stabilizing.The ‹en› bigram shows no clear proportionality between aspect ratio and inter-letter distance, while ‹de› has the smallest distance across all profiles and exhibits rare line-final spacing variation in GP1.
5 Conclusion
The paper presents a framework that unifies morphological and metrological paleographic analysis through detection-based recognition and prototype-based reconstruction. In the codex case study, its visualized measurements support hand separation and reveal subtle paleographic phenomena.
- The framework combines detection-based text recognition with prototype-based image reconstruction to produce stable, efficient training and higher-quality character prototypes.
- Measurements and visualizations of proportionality and horizontal compression provide criteria for hand separation and uncover subtle paleographic phenomena.