Kurdish Scene Text Recognition Version 2.0 (KSTRV2) Dataset

KSTRV2 (Kurdish STR) Version 2 

KSTRV2 is a large-scale dataset developed for Kurdish Scene Text Recognition (KSTR), addressing the scarcity of resources for non-Latin script like Kurdish. It includes 1,420 real-world scene images and 19,872 extracted word-level samples across Kurdish (Sorani and Badini dialects), Arabic, and English. To expand coverage and improve generalizability, the dataset is augmented with 20,000 synthetic text examples, crafted with diverse typography, multi-angle orientations, simulated distortions, and intricate background textures. This synthesis enhances the dataset’s capacity to handle real-world variability, supporting robust training for text recognition systems in underrepresented languages.

KSTRV2 — Release Notes (September 2025)

Overview

KSTRV2 is an updated release of the Kurdish Scene Text Recognition dataset. It standardizes text encodings, fixes minor annotation issues, and rebalances a small slice of synthetic samples to better reflect realistic difficulty. Use KSTRV2 for any new training or benchmarking; results based on KSTRV1 are not directly comparable.

Dataset record (DOI): https://doi.org/10.5281/zenodo.15038953 (Zenodo) zenodo.org

Article describing KSTRV1: “KSTRV1: A scene text recognition dataset for central Kurdish (Arabic-Based) script,” Data in Brief, 2025. https://www.sciencedirect.com/science/article/pii/S2352340925003786 

Github Link : link

What’s new in v2.0

Unified Unicode encoding for annotationsSome visually identical characters had different code points due to mixed dictionary sources. All annotations are now standardized to a single encoding scheme to remove ambiguity and improve model robustness and reproducibility.

 Rebalanced synthetic subset difficultyA small subset of synthetic word images that was overly complex has been replaced with easy-to-medium complexity images. The goal is to keep diversity while making the training and eval difficulty more representative of real scenes.

Annotation quality fixesMinor spelling and labeling errors identified during validation were corrected.

Release readinessWith these changes, KSTRV2 is cleaner and more consistent, ready for both model development and standardized benchmarking.

RecommendationPlease migrate to KSTRV2 for all new work; KSTRV1 contains known inconsistencies that are resolved here.

Compatibility & migration notes

Do not mix versions in training or evaluation. Report results on KSTRV1 and KSTRV2 separately to avoid confounding from the encoding and labeling changes.

Unicode handling: ensure your preprocessing preserves the dataset’s normalized encoding; disable any unintended re-normalization that could remap code points and break label alignment. (Change motivated by the v2 standardization.)

Synthetic curriculum: if you used difficulty-based sampling on v1, re-tune sampling weights given the rebalanced synthetic subset.

Dataset scope (context)

KSTR includes 1,420 natural scene images and 19,872 cropped word samples across Kurdish (Sorani and Badini), Arabic, and English, plus 20,000 synthetic instances—details in the dataset record and paper. zenodo.org 

How to cite

 Dataset (specific version once v2 is public):Kurdish Scene Text Recognition (KSTR), Version 2.0. Zenodo. DOI: 10.5281/zenodo.15038953 (link resolves to the latest version; please use the versioned DOI shown on the Zenodo “Versions” panel once v2 is live).

Article:Salih, S. O., & Jacksi, K. (2025). KSTRV1: A scene text recognition dataset for central Kurdish (Arabic-Based) script. Data in Brief.

License

CC BY 4.0. You may reuse and redistribute with attribution. zenodo.org

Changelog (Keep-a-Changelog style)

Changed: Standardized Unicode encoding across all annotations.

Changed: Rebalanced a small synthetic subset to easy/medium difficulty.

Fixed: Minor spelling and labeling errors.

Note: KSTRV1 is deprecated for new experiments.

KSTRV2（库尔德语场景文本识别（Kurdish Scene Text Recognition, KSTR）版本2）

KSTRV2是专为库尔德语场景文本识别任务开发的大规模数据集，旨在填补库尔德语等非拉丁脚本场景文本识别的资源空白。该数据集包含1420张真实场景图像，以及覆盖库尔德语（索拉尼方言与巴迪尼方言）、阿拉伯语和英语的19872个提取得到的单词级样本。为扩大覆盖范围并提升模型泛化能力，数据集还扩充了20000个合成文本样本，此类样本采用多样化排版、多角度朝向、模拟失真与复杂背景纹理生成。此类合成操作可增强数据集应对真实场景变异性的能力，为代表性不足语言的文本识别系统提供稳健的训练支撑。

KSTRV2——2025年9月发布说明

## 概述
KSTRV2是库尔德语场景文本识别数据集的更新版本。其统一了文本编码格式，修复了少量标注问题，并对一小部分合成样本进行了重平衡，使其更贴合真实场景的难度分布。所有新的训练或基准测试均应使用KSTRV2；基于KSTRV1的结果无法直接与本版本进行对比。

数据集记录（数字对象标识符（Digital Object Identifier, DOI））：https://doi.org/10.5281/zenodo.15038953（Zenodo平台）
介绍KSTRV1的论文：《KSTRV1：面向中库尔德语（基于阿拉伯字母）脚本的场景文本识别数据集》，《数据简报》，2025年。https://www.sciencedirect.com/science/article/pii/S2352340925003786
GitHub链接：链接

## v2.0版本更新内容
1. **统一标注的统一码（Unicode）编码**：此前由于混合使用不同字典来源，部分视觉上完全相同的字符拥有不同的码点。目前所有标注均已统一为单一编码方案，消除歧义并提升模型鲁棒性与可复现性。
2. **重平衡合成子集的难度**：将一小部分过于复杂的合成单词图像替换为易至中等难度的图像，目标是在保留样本多样性的同时，让训练与评估难度更贴合真实场景的实际情况。
3. **标注质量修复**：修正了验证过程中发现的少量拼写与标注错误。
4. **发布就绪性**：经过上述改进，KSTRV2更为简洁统一，可用于模型开发与标准化基准测试。
5. **使用建议**：所有新研究均应迁移至KSTRV2；KSTRV1存在已知的不一致问题，本版本已完全解决。

## 兼容性与迁移说明
- 请勿在训练或评估过程中混合使用两个版本。请分别报告基于KSTRV1与KSTRV2的实验结果，避免因编码与标注变更产生结果混淆。
- 统一码（Unicode）处理：确保预处理步骤保留数据集的归一化编码格式，禁用任何可能重新映射码点、破坏标签对齐的意外重归一化操作（此变更源于v2版本的标准化需求）。
- 合成样本课程学习：若你曾在v1版本中使用基于难度的合成课程学习采样方式，请针对重平衡后的合成子集重新调整采样权重。

## 数据集范围（背景说明）
KSTR包含1420张自然场景图像与19872个裁剪后的单词样本，覆盖库尔德语（索拉尼与巴迪尼方言）、阿拉伯语与英语，外加20000个合成样本——详细信息请参见数据集记录与对应论文。（Zenodo平台）

## 引用方式
- 数据集（v2正式公开后请使用指定版本）：库尔德语场景文本识别（KSTR），版本2.0。Zenodo。DOI: 10.5281/zenodo.15038953（该链接指向最新版本；v2正式上线后，请使用Zenodo“版本”面板中显示的带版本号的DOI）。
- 论文：Salih, S. O., & Jacksi, K. (2025). KSTRV1: A scene text recognition dataset for central Kurdish (Arabic-Based) script. *Data in Brief*.

## 许可协议
CC BY 4.0。您可在注明原作者出处的前提下，重用或重新分发该数据集。（Zenodo平台）

## 变更日志（遵循Keep-a-Changelog格式）
- 变更：统一所有标注的统一码（Unicode）编码。
- 变更：将一小部分合成子集重平衡为易/中等难度。
- 修复：修正少量拼写与标注错误。
- 备注：KSTRV1已不适合用于新实验。