NUST-UHWR: An Unconstrained Offline Urdu Handwriting Recognition Dataset with Line Images and Ground Truth Transcriptions

Dataset Overview NUST-UHWR is an unconstrained offline Urdu handwriting recognition dataset developed for research in handwritten text recognition, optical character recognition, document image analysis, computer vision and low resource language processing. The dataset contains scanned images of handwritten Urdu text lines together with their corresponding ground truth transcriptions. The dataset was created to address the limited availability of large and diverse resources for Urdu handwriting recognition. Urdu is commonly written in the Nastaleeq script, which presents significant recognition challenges because of its cursive structure, context dependent character shapes, vertical character overlap, varying ligature formations and the visual similarity of several characters. The original study describes NUST-UHWR as the first unconstrained Urdu handwriting dataset compiled from approximately 1,000 people representing different backgrounds, age groups and genders. The text material was selected from seven different fields to provide coverage of frequently used words from multiple domains. Dataset Purpose The dataset is intended to support the development, training, validation and evaluation of systems for offline Urdu handwritten text recognition. It can be used for segmentation free recognition, sequence modelling, character level transcription, language modelling and benchmarking of machine learning and deep learning architectures. NUST-UHWR is particularly suitable for evaluating models that must recognise previously unseen handwriting styles. It contains natural variation in writing shape, stroke formation, spacing, character connection, ligature structure and line width. The dataset can support CNN, CNN-RNN, CTC, encoder decoder, attention based and Transformer based recognition systems. It may also be used for multilingual transfer learning involving Urdu and other Arabic script languages. Dataset Composition The uploaded archive contains 10,606 grayscale JPEG images of handwritten Urdu text lines. All image files have a height of 64 pixels. Their widths vary according to the length of the handwritten line and range from approximately 200 to 2,170 pixels in the uploaded archive. The archive contains the following principal components: A directory containing the handwritten line images in JPEG format A training manifest named train.txt A validation manifest named val.txt A testing manifest named test.txt A character vocabulary file named chars.txt A legacy character vocabulary file named old_chars.txt A character frequency file named freq.txt The chars.txt file contains 116 unique symbols that occur in the supplied ground truth transcriptions. These include Urdu letters, Urdu numerals, selected Arabic script symbols, punctuation marks and a limited number of Latin characters and digits. Ground Truth and Annotation Format Every labelled sample consists of a handwritten line image and its corresponding Unicode text transcription. The training, validation and testing manifest files use a tab separated format. Each record contains: The relative path of the image file A tab character The complete Urdu ground truth transcription The transcription files should be read with UTF-8 encoding to preserve Urdu characters, diacritics, numerals and punctuation correctly. The dataset is structured for line-level handwritten text recognition. It is not a character segmentation dataset and does not provide individual bounding boxes for characters, words or ligatures. Recognition systems are expected to learn the alignment between the image features and the output text sequence. Recommended Data Splits The original publication reports the following official distribution: Training samples: 8,484 Validation samples: 1,061 Testing samples: 1,061 Total samples: 10,606 The later Conv-Transformer paper uses the same official dataset statistics. The uploaded archive currently contains: Training manifest records: 8,483 Validation manifest records: 1,061 Testing manifest records: 1,061 Total labelled manifest records: 10,605 The image directory contains 10,606 JPEG files. One image named 03598.jpg is not referenced in the training, validation or testing manifest. This appears to account for the difference between the published total and the supplied manifest total. The image does not have an identifiable transcription in the provided split files. This inconsistency should be resolved before the final Zenodo release. The missing transcription may be added to the appropriate manifest if it can be recovered from the authoritative source. Otherwise, the unreferenced image should be documented clearly or removed from the deposited version. Data Collection and Diversity The dataset was compiled from approximately 1,000 writers. The original paper describes the contributors as representing different ages, genders and backgrounds. This diversity was included to capture natural differences in handwriting styles and to improve the ability of recognition models to generalise to unseen writers. The textual content was selected from seven fields to provide vocabulary from multiple domains and to increase the representation of commonly used Urdu words. The source material was designed to produce varied sentence structures, character combinations and ligature formations. The deposited files do not include writer identifiers, demographic labels or a detailed participant level metadata table. The dataset should therefore be treated as a handwriting recognition corpus rather than a demographic or writer identification dataset. Image Characteristics The images are stored as grayscale JPEG files. Each image contains a single handwritten Urdu text line. The fixed image height supports consistent model input preparation while the variable width preserves differences in sentence length and handwriting spacing. The images contain natural variations in stroke thickness, writing alignment, word spacing, character overlap, calligraphic complexity and image quality. The associated research paper describes model preprocessing in which images were converted to grayscale and resized to a fixed height of 64 pixels while maintaining the aspect ratio. Width padding was then used during model batching. Recommended Research Applications The dataset may be used for: Urdu handwritten text recognition Offline handwriting recognition Optical character recognition Line-level sequence recognition Segmentation free text recognition Urdu language modelling Character error rate benchmarking CNN-RNN and CTC model evaluation Transformer based document recognition Multilingual and cross script transfer learning Low resource language technology Document digitisation and information extraction Recognition of Arabic derived cursive scripts Investigation of handwriting variation and recognition robustness Quality Considerations The dataset provides a valuable benchmark for unconstrained Urdu handwriting recognition. However, users should consider the natural difficulties associated with Urdu handwriting, including overlapping characters, ambiguous shapes, varying ligature structures, inconsistent spacing and complex calligraphic writing. The later Conv-Transformer study documents examples of label noise, distorted input images and cases in which the supplied ground truth contains additional or incorrect text. It also shows that some model predictions appear correct even where the provided ground truth is mismatched. Researchers are advised to inspect uncertain samples during error analysis and to distinguish genuine model errors from annotation inconsistencies. Limitations NUST-UHWR contains line-level images and does not provide full pages, paragraphs, word boxes, character boxes or writer identity labels. The exact demographic distribution of contributors is not included in the deposited archive. The seven source domains are described in the publication, but domain labels are not attached to individual samples in the uploaded version. Some ground truth records may contain spelling variation, punctuation inconsistencies, annotation noise or characters that are visually unclear. The current archive has one image that is not referenced in any split manifest. This should be corrected or documented before public release. The dataset should not be used to identify individual writers or to infer personal characteristics from handwriting. Reproducibility The predefined training, validation and testing files should be retained to support direct comparison with published results. Researchers should report whether they used the supplied split or created a different partition. Evaluation should preferably include Character Error Rate. Any normalisation of spaces, punctuation, diacritics, Unicode characters or numerals should be reported clearly. When preprocessing the images, researchers should document the target height, width handling, padding, resizing method, image normalisation and augmentation procedures. Associated Publications The dataset was originally introduced in: Noor ul Sehr Zia, Muhammad Ferjad Naeem, Syed Muhammad Kumail Raza, Muhammad Mubasher Khan, Adnan Ul-Hasan and Faisal Shafait. A convolutional recursive deep architecture for unconstrained Urdu handwriting recognition. Neural Computing and Applications, Volume 34, pages 1635 to 1648. DOI: 10.1007/s00521-021-06498-2 Published online: 14 September 2021 The dataset was later used in: Nauman Riaz, Haziq Arbab, Arooba Maqsood, Khuzaeymah Nasir, Adnan Ul-Hasan and Faisal Shafait. Conv-transformer architecture for unconstrained off-line Urdu handwriting recognition. International Journal on Document Analysis and Recognition, Volume 25, pages 373 to 384. DOI: 10.1007/s10032-022-00416-5 Published online: 23 September 2022 The later study trained a Conv-Transformer architecture on NUST-UHWR and reported a test Character Error Rate of 5.31 percent when handwritten and printed Urdu datasets were combined during training.