Datasets:
File size: 2,335 Bytes
eccb364 dd9a802 cf954f2 dd9a802 cf954f2 9ddc5fc cf954f2 dd9a802 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 | ---
license: mit
language:
- fa
tags:
- HTR
- Arabic-scripts
- persian-HTR
size_categories:
- 1K<n<10K
---
# PHTD Line-Level Dataset (Cleaned and Split Version)
**Important:** I am *not* the creator or copyright holder of the original PHTD dataset.
The underlying handwritten Persian page images and pixel-level masks were introduced in the following works:
1. **Alaei et al., “A New Dataset of Persian Handwritten Documents and Its Segmentation,” **
2. **Alaei, Pal & Nagabhushan, “Dataset and ground truth for handwritten text in four different scripts,” **
This repository provides a **processed, line-level version** of that dataset, created for reproducible handwritten text recognition (HTR) research and for use with the CRHV framework.
---
## 📌 Citation
If you use this processed line-level dataset, please cite our paper:[CER-HV: A CER-Based Human-in-the-Loop Framework for Cleaning Datasets Applied to Arabic-Script HTR](https://www.arxiv.org/abs/2601.16713)
and cite the original creators
## ✨ What This Version Provides
The original PHTD dataset contains **page-level images** with **pixel-wise segmentation masks** identifying individual text lines.
However, it does **not** include:
- extracted line images,
- standardized train/validation/test splits,
- or a leakage-free partition.
This dataset aims to provide exactly that.
### ✔ Line Image Extraction
Using the original pixel masks provided in the dataset, each text line was isolated by:
- extracting the minimal bounding box of each mask region,
- applying a 5-pixel padding margin,
- masking out all non-target pixels,
- and generating a clean cropped line image.
This results in accurate line-level samples that preserve the original handwritten content.
### ✔ Leakage-Free Dataset Splits
The original page set contains **near-duplicate pages** and text overlaps.
To prevent data leakage between training and evaluation splits:
- we computed pairwise similarity between pages,
- identified overlapping and duplicate pages,
- restricted validation and test sets to *non-overlapping* pages only.
This ensures that evaluation is fair and does not unintentionally benefit from training-page content.
### ✔ Standardized Train/Validation/Test Structure
Each split follows a unified directory structure: |