metadata

language:
  - fa
pretty_name: Persian Historical Documents Handwritten Characters
size_categories:
  - 1K<n<10K
tags:
  - ocr
  - character-recognition
  - persian
  - historical
  - handwritten
  - nastaliq
  - character

Persian Historical Documents Handwritten Characters

Dataset Description

Repository: https://github.com/iarata/persian-docs-ocr
Paper: https://doi.org/10.1007/978-3-031-53969-5_20
Point of Contact: hajebrahimi.research [at] gmail [dot] com

Summary

This dataset contains pre-processed images of Persian characters' contextual forms (except letter گ) from 5 handwritten Persian historical books written in Nastaliq script. The dataset contains 2775 images of 111 classes. The images are in TIFF format and have a resolution of 72 dpi. The images are in black and white and have a size of 395 × 395 pixels.

Languages

Persian

Dataset Structure

The dataset is structured as follows:

├── data
│   ├── 06a9_01.tif
│   ├── 06a9_02.tif
│   ├── 06a9_03.tif
│   ├── 06a9_04.tif
│   ├── 06a9_05.tif
│   ├── ...
│   ├── 06a9_25.tif
│   │
│   ├── 06cc_01.tif
│   ├── 06cc_02.tif
│   ├── 06cc_03.tif
│   ├── 06cc_04.tif
│   ├── 06cc_05.tif
│   ├── ...
│   ├── 06cc_25.tif
│   ├── ...

The naming of each image indicates the UTF-16 hexadecimal code (Hex to String Decoder) of a character's contextual form followed by the number of the image. In the numbering, every 5 images are from a new book. The contextual form of every character is treated as a separate class resulting in 111 classes.

Dataset Creation

For building this dataset 5 historical Persian books from the Library of Congress

Source Data

The data was collected from 5 historical Persian books from the Library of Congress. The books are as follows:

The images were pre-processed using the following steps:

Images were first normalized to reduce noise from the background of the characters. The normalized image is then converted to a single-channel grayscale image. Following that, image thresholding is applied to the grayscale image to remove the characters' background. The thresholded image is binarized so that the pixel values greater than 0 become 255 (white), and pixels with a value of 0 (black) remain unchanged. Finally, the binarized image is inversed.

Annotations

Before pre-processing the images the characters were cropped from the books and were saved with their UTF-16 hexadecimal code plus the number of the image (e.g. 06a9_01.tif).

Annotators:

Citation Information

Hajebrahimi, A., Santoso, M.E., Kovacs, M., Kryssanov, V.V. (2024). Few-Shot Learning for Character Recognition in Persian Historical Documents. In: Nicosia, G., Ojha, V., La Malfa, E., La Malfa, G., Pardalos, P.M., Umeton, R. (eds) Machine Learning, Optimization, and Data Science. LOD 2023. Lecture Notes in Computer Science, vol 14505. Springer, Cham. https://doi.org/10.1007/978-3-031-53969-5_20

BibTeX:

@InProceedings{10.1007/978-3-031-53969-5_20,
    author="Hajebrahimi, Alireza
    and Santoso, Michael Evan
    and Kovacs, Mate
    and Kryssanov, Victor V.",
    editor="Nicosia, Giuseppe
    and Ojha, Varun
    and La Malfa, Emanuele
    and La Malfa, Gabriele
    and Pardalos, Panos M.
    and Umeton, Renato",
    title="Few-Shot Learning for Character Recognition in Persian Historical Documents",
    booktitle="Machine Learning, Optimization, and Data Science",
    year="2024",
    publisher="Springer Nature Switzerland",
    address="Cham",
    pages="259--273",
    abstract="Digitizing historical documents is crucial for the preservation of cultural heritage. The digitization of documents written in Perso-Arabic scripts, however, presents multiple challenges. The Nastaliq calligraphy can be difficult to read even for a native speaker, and the four contextual forms of alphabet letters pose a complex task to current optical character recognition systems. To address these challenges, the presented study develops an approach for character recognition in Persian historical documents using few-shot learning with Siamese Neural Networks. A small, novel dataset is created from Persian historical documents for training and testing purposes. Experiments on the dataset resulted in a 94.75{\%} testing accuracy for the few-shot learning task, and a 67{\%} character recognition accuracy was observed on unseen documents for 111 distinct character classes.",
    isbn="978-3-031-53969-5"
}