Uncovering the Handwritten Text in the Margins: End-to-end Handwritten Text Detection and Recognition

Paper Info

Title	Uncovering the Handwritten Text in the Margins: End-to-end Handwritten Text Detection and Recognition
Authors	Liang Cheng, Jonas Frankemölle, Adam Axelsson, Ekta Vats
Conference	Proceedings of the 8th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature (LaTeCH-CLfL 2024)
Paper	aclanthology: 2024.latechclfl-1.12
Code	GitHub: Project-Marginalia

---

Motivation

Libraries and archives worldwide are digitizing their historical collections at a massive scale. The printed text in early books is well-documented and searchable. But what about the handwritten notes in the margins — the marginalia?

Marginalia are a precious historical resource:

• They reveal how earlier readers engaged with texts
• They provide insights into intellectual history, gender studies, and knowledge exploration
• Collections like the Walleriana at Uppsala University Library (dating back to the 1400s) contain rich marginalia that have never been digitally transcribed

Challenges:

1. High variability — different writers, centuries, languages, scripts
2. Data scarcity — expert annotation is expensive and time-consuming; only 513 labeled pages in our dataset
3. Under-explored problem — most HTR research focuses on main text, not marginalia

---

Pipeline Overview

Scanned book page (350×500)
        │
        ▼
┌──────────────────────────┐
│ Marginalia Detection      │
│ (Faster R-CNN / R-CNN)    │ ← outputs bounding boxes
└────────────┬─────────────┘
             │ cropped marginalia regions
             ▼
┌──────────────────────────┐
│ Segmentation              │
│  • Sobel + horizontal     │ ← line segmentation
│    projection             │
│  • Binarization + vertical│ ← word segmentation
│    projection             │
└────────────┬─────────────┘
             │ individual word images
             ▼
┌──────────────────────────┐
│ AttentionHTR              │
│  TPS → ResNet32 → BiLSTM  │ ← text transcription
│  → attention decoder      │
└──────────────────────────┘
             │
             ▼
    Optional: Language model
    post-processing (BERT)

---

Method Details

1. Data & Preprocessing

Item	Details
Source	Uppsala University Library early book collections (Walleriana)
Labeled images	513 pages, expert-annotated with LabelMe (bounding boxes only)
Train/test split	9:1
Data augmentation	Horizontal flip + Gaussian noise + random brightness/contrast → 1848 training samples
Image size	Resized to 350×500

2. Marginalia Detection: R-CNN vs Faster R-CNN

R-CNN baseline:

• Region proposals via MSER (Maximally Stable Extremal Regions) — a watershed-based algorithm
• Positive samples: crop from annotated boxes, resize to 227×227
• Negative samples: 4 MSER boxes with IoU=0 per image
• Classifier: AlexNet (5 conv + 2 FC layers, 2 output classes)
• Problem: depends entirely on MSER quality → fails when MSER doesn’t capture full marginalia

Faster R-CNN (our main method):

• Backbone: ResNet-50 (pre-trained)
• RPN (Region Proposal Network) replaces selective search/MSER
• End-to-end training: RPN and Fast R-CNN share convolutional features
• Output: 2 classes (marginalia / non-marginalia) + 4 bounding box coordinates
• Training: 13 epochs, learning rate 0.001

3. Segmentation

Line segmentation:

1. Sobel edge detection to emphasize text strokes
2. Horizontal projection (sum of pixel values per row)
3. Peak detection to find line positions
4. Threshold = midpoint between highest peak and lowest value

Word segmentation:

1. Binarize each line image
2. Vertical projection to find spaces
3. Calculate average space width
4. Split only on spaces wider than the average

4. Recognition: AttentionHTR

We reuse our previously developed AttentionHTR model (Kass & Vats, 2022):

Stage	Description
TPS Transform	Thin-plate spline transformation to correct text distortion
ResNet-32	32-layer ResNet for feature extraction
2-layer BiLSTM	Bidirectional LSTM for sequence modeling
Attention + Decoder	Content-based attention mechanism with unidirectional LSTM decoder

Key advantage: Pre-trained on Imgur5K (5000 different writers), so it transfers well to our small dataset with no additional training needed.

---

Results

Detection Performance

Model	Performance	IoU	Key Failure Modes
R-CNN	~90.7% val accuracy	Low	(1) Boxes don’t fully align (2) Only partial marginalia captured (3) Fails on large marginalia
Faster R-CNN	Accurate across diverse cases	0.82	Robust to large marginalia, figures, and varying layouts

Why Faster R-CNN wins: It doesn’t depend on pre-segmentation (MSER). The RPN learns to propose relevant regions directly from the image feature map, and the shared convolutional backbone makes it both faster and more accurate.

Segmentation Performance

• Works well when: lines don’t intersect, words are clearly separated
• Fails when:
  • Handwriting rows overlap (letters from two lines share the same pixel rows) → whole block detected as one line
  • Words are connected by strokes or underlines → can’t separate words
• Root cause: Projection-based methods are purely heuristic and can’t handle complex layouts

Recognition Performance

Input quality	AttentionHTR performance
Clear, well-segmented single word	Very accurate — words and numbers both recognized correctly
Blurry/unreadable (even for humans)	Fails — as expected
Multiple words in one image	Inaccurate

Key finding: Recognition accuracy is bounded by segmentation quality. The better the input, the better the output — and AttentionHTR requires zero fine-tuning on our data to work well.

---

Discussion & Limitations

Aspect	Discussion
Detection bottleneck	R-CNN’s dependence on MSER is its Achilles’ heel; Faster R-CNN solves this cleanly
Segmentation bottleneck	Projection-based segmentation is the weakest link in the pipeline
Future: better segmentation	Could use R-CNN at the word level too, but requires word-level annotation data
Other detectors	Mask R-CNN and YOLO not yet explored; data format is compatible → future work
Other recognizers	TrOCR (Transformer-based) not compared; we chose AttentionHTR for its strong pre-trained multi-writer model
Language model	Skip-gram or BERT post-processing could improve accuracy but adds compute cost
Data limitation	Only bounding box labels, no text annotations → can’t fine-tune the recognizer on our data
Generalization	Special scripts (Gothic, curlicue), blurry text, strike-throughs, very poor handwriting remain challenging

---

Key Takeaways

1. First end-to-end marginalia detection + recognition pipeline — to our knowledge, the first work addressing historical handwritten marginalia as a complete pipeline
2. Faster R-CNN is the right tool for detection — IoU 0.82, substantially better than MSER+R-CNN
3. Segmentation is the pipeline bottleneck — heuristic projection methods limit overall accuracy
4. Transfer learning makes it work with scarce data — pre-trained AttentionHTR handles the recognition side without additional training
5. Reproducible and extensible — modular code, adaptable data format, ready for future improvements (Mask R-CNN, YOLO, TrOCR, language models)

One-sentence summary: An end-to-end pipeline combining Faster R-CNN detection, projection-based segmentation, and transfer-learned attention HTR enables automatic extraction of handwritten marginalia from historical books, even with limited training data.

---

BibTeX

@inproceedings{cheng2024uncovering,
  title={Uncovering the Handwritten Text in the Margins: End-to-end Handwritten Text Detection and Recognition},
  author={Cheng, Liang and Frankem{\"o}lle, Jonas and Axelsson, Adam and Vats, Ekta},
  booktitle={Proceedings of the 8th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature (LaTeCH-CLfL)},
  pages={111--120},
  year={2024}
}