wahlinski/handwritten_cross-outs

HTR with Cross-out Words Dataset

Overview

This dataset consists of handwritten word images produced by 12 different authors. It includes both clean (non-crossed-out) samples and crossed-out words, making it suitable for multiple handwriting-related research tasks.

The dataset introduces 7 distinct cross-out types, along with a mixed subset that combines clean samples and randomly selected examples from different cross-out categories. This design supports both binary and multi-class classification scenarios.

Each word image is annotated with:

• Cross-out label (label): one of "clean", "single-line", "double-line", "diagonal", "cross", "wave", "zig-zag", "scratch", or "mix"
• Writing tool (writing_tool): one of "ballpoint", "digital_black", "digital_gray", "marker", or "pencil"
• Transcription (text): ground truth text

---

Dataset Structure

Each sample in the dataset contains:

• image: Handwritten word image
• label: Label indicating clean or cross-out type
• writing_tool: Writing instrument used
• text: Ground truth transcription

The dataset is structured as a multi-task dataset, enabling flexible use across different tasks.

Usage

The sample code below demonstrates how to download the dataset and load the first image from the training set.

from datasets import load_dataset
from PIL import Image
import matplotlib.pyplot as plt

dataset = load_dataset("wahlinski/handwritten_cross-outs")
print(dataset)

sample = dataset["train"][0]
print(sample)

plt.imshow(sample['image'])
plt.axis("off")
plt.title(sample["label"])
plt.show()

---

Cross-out Categories

The dataset includes the following categories:

• 7 distinct cross-out types: single-line, double-line, diagonal, cross, wave, zig-zag, scratch
• Clean: non-crossed-out samples
• Mixed: combination of clean and randomly sampled cross-out instances

---

Tasks and Use Cases

This dataset can be used for:

1. Binary Classification

• Clean vs. Crossed-out (e.g., using clean and mixed subsets)

2. Multi-class Classification

• Classification of the 7 cross-out types

3. Writing Tool Classification

• Predicting the writing instrument used

4. Handwritten Text Recognition (HTR)

• Training models using the provided transcriptions

---

Research Context

This dataset is introduced in the paper:

"A Study of Handwritten Text Recognition with Cross-out Words"

If you use this dataset in your research, please cite the following:

Citation details will be added upon publication.

---

Notes

• The dataset includes variability across writers and writing tools.
• Cross-out variations introduce realistic challenges for document analysis systems.
• Suitable for both academic research and benchmarking.

---

Contact

For questions or collaboration, please contact:

• Filip Wåhlin