Urdu Katib Handwritten Dataset: A Historical Document Dataset for Offline Urdu Handwritten Text Recognition with CRNN-Based Baseline Evaluation

2026-06-18 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Natural Language Processing · Depth: Advanced, extended

Summary

The Urdu Katib Handwritten Dataset (UKHD) is introduced as the first offline Urdu handwritten text lines dataset specifically curated from historical documents written by Katibs in the Nastalique calligraphic style. This dataset addresses the scarcity of benchmark resources for Urdu Handwritten Text Recognition (UHTR), a challenging task due to Urdu's cursive, diagonal, overlapping, and context-sensitive script. The study also evaluates CRNN-based hybrid models on UKHD, identifying the CNN-BGRU-CTC model as the most robust performer. This model achieved average Character Error Rate (CER) of 5.2% and Word Error Rate (WER) of 16.9% on the test set, demonstrating strong performance in recognizing historical Urdu script and its unique diacritics.

Key takeaway

For AI Scientists and Machine Learning Engineers working on historical document preservation, this research highlights the critical need for specialized datasets like UKHD. You should prioritize CRNN-based models, specifically the CNN-BGRU-CTC architecture, for Urdu Handwritten Text Recognition, as it demonstrated superior performance (5.2% CER). Consider integrating advanced image enhancement and transformer-based post-processing to further improve recognition rates for challenging cursive scripts.

Key insights

The Urdu Katib Handwritten Dataset (UKHD) and CRNN-BGRU-CTC model establish a baseline for historical Urdu handwritten text recognition.

Principles

• Cursive scripts like Urdu pose unique HTR challenges due to diagonal, overlapping, and context-sensitive characters.
• Bidirectional RNN variants (BLSTM, BGRU) are superior for image-based sequence recognition in cursive scripts.
• Implicit segmentation-based recognition with hybrid deep learning models is effective for UHTR.

Method

UKHD generation uses semi-automatic image acquisition, preprocessing (skew correction via HPP), line segmentation (HPP-based, manual adjustment), and annotation (Cloud Vision API with manual correction). CRNN models combine CNN feature extraction, RNN sequence modeling, and CTC for alignment-free transcription.

In practice

• Utilize the UKHD dataset for developing robust UHTR systems for historical Urdu literature.
• Implement CNN-BGRU-CTC architecture for optimal performance in Urdu Katib Handwriting Recognition (UKHR).
• Employ semi-automatic methods for efficient dataset creation, combining automated transcription with manual review.

Topics

• Urdu Handwritten Text Recognition
• Urdu Katib Handwritten Dataset
• CRNN Models
• Nastalique Calligraphy
• Historical Document Preservation
• Deep Learning for OCR

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Urdu Katib Handwritten Dataset: A Historical Document Dataset for Offline Urdu Handwritten Text Recognition with CRNN-Based Baseline Evaluation — The Urdu Katib Handwritten Dataset (UKHD) addresses a critical gap in resources for offline Urdu Handwritten Text Recognition.
• Performance Gap Analysis between Latin and Arabic Scripts HTR — The HTR performance gap between Latin and Arabic scripts persists due to higher visual variability and character similarity, even with extensive data.
• Why Nepali OCR Is Brutally Hard — And How We Finally Solved It — A hybrid deep learning pipeline overcomes complex script challenges to automate document data extraction and form pre-filling.
• myMNIST: Benchmark of PETNN, KAN, and Classical Deep Learning Models for Burmese Handwritten Digit Recognition — CNNs remain a strong baseline for Burmese handwritten digit recognition, with PETNNs showing competitive performance.
• Does anyone have a copy of the ICDAR2013 Chinese Handwriting Competition Dataset? [R] — The ICDAR2013 Chinese Handwriting Dataset is difficult to obtain, but a partial LMDB version exists.
• Text region detection in historical astronomical diagrams — A new large-scale dataset and Poly-DETR baseline advance text detection in historical astronomical diagrams.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.