Yaredoffice/geez-characters

Amharic (Geʽez) Handwritten Character Dataset (32×32)

Dataset Details

Description

This dataset contains handwritten images of Amharic (Geʽez script) characters intended for character-level Optical Character Recognition (OCR) and handwriting recognition research.

Property	Value
Total Images	13,000+
Classes	287 distinct characters
Image Size	32 × 32 pixels
Format	Grayscale
Distribution	Balanced across all classes

The dataset is designed to support CPU-efficient character classifiers and low-resource language research, particularly for Ethiopic scripts.

• Curated by: Yared
• Language: Amharic (Geʽez / Ethiopic script)
• License: Apache License 2.0

---

Uses

Direct Use

• Training and evaluating handwritten character classifiers
• OCR pipelines that operate at character level
• Research on low-resource and underrepresented scripts
• Benchmarking lightweight CNN models on constrained hardware (CPU, low RAM)

Out-of-Scope Use

• Writer identification or biometric analysis
• Forensic handwriting attribution
• Recognition of printed or typeset Amharic text
• Word-level or sentence-level language modeling without additional segmentation

---

Dataset Structure

Data Fields

Each sample contains:

Field	Description
image	32×32 grayscale image of a single handwritten character
label	Integer class index in range [0, 286]

Directory Layout

dataset/
├── train/
│   ├── 0/
│   ├── 1/
│   ├── ...
│   └── 286/
└── test/
    ├── 0/
    ├── 1/
    ├── ...
    └── n/

Folder names correspond directly to character class IDs.

---

Dataset Creation

Curation Rationale

Publicly available datasets for handwritten Ethiopic scripts are scarce, especially at character level. This dataset was created to provide a standardized, balanced, and lightweight benchmark for Amharic handwritten character recognition, enabling both academic research and practical OCR system development under limited computational resources.

Source Data

Data Collection and Processing

1. Handwritten characters were collected on paper forms
2. Pages were scanned or photographed
3. Individual characters were extracted and cropped
4. Images were converted to grayscale
5. Resized to a fixed resolution of 32×32 pixels
6. Manually organized into class-specific directories

No synthetic data generation was used.

Source Data Producers

The handwritten samples were produced by human contributors mainly in an academic native environment though a portion of participants are also tigrinya native. No personally identifiable information is associated with the samples.

---

Annotations

Annotation Process

Annotations are implicit and directory-based. Each image inherits its label from the directory name representing a specific Geʽez character class. This mapping serves as the ground-truth annotation.

Annotators

Annotation and class assignment were performed by the dataset creator during dataset organization and validation.

Personal and Sensitive Information

This dataset does not contain:

• Names or identifiers
• Demographic metadata
• Sensitive personal information

The dataset consists solely of isolated handwritten character images.

---

Bias, Risks, and Limitations

Consideration	Description
Demographic bias	Handwriting styles may reflect a limited demographic group due to localized data collection from less than 500 Dire Dawa Universty Students only
Style coverage	Extreme handwriting variations (e.g., elderly or non-academic writers) may be underrepresented
Scope limitation	Character-level only; does not capture word or sentence context and due to the 500 participants some unique paterns might not be collected

Recommendations

• Fine-tune models with additional local handwriting samples for deployment
• Combine this dataset with document-level segmentation pipelines when building full OCR systems
• Apply data augmentation to improve robustness to handwriting variability

---

Citation

If you use this dataset in your work, please cite it as follows:

BibTeX

@dataset{amharic_handwritten_characters_2024,
  author       = {Yared},
  title        = {Amharic (Geʽez) Handwritten Character Dataset},
  year         = {2024},
  publisher    = {Hugging Face},
  license      = {Apache-2.0},
  url          = {https://huggingface.co/datasets/Yaredoffice/geez-characters}
}

APA

Yared. (2024). Amharic (Geʽez) Handwritten Character Dataset. Hugging Face. https://huggingface.co/datasets/Yaredoffice/geez-characters

---

Dataset Card Authors

Yared

Contact

For questions or contributions, please reach out via the dataset's Hugging Face discussion tab or the author's GitHub profile.