Urdu Handwriting Recognition with Deep Learning: Current Methods and Future Prospects

Abstract

The Nastaliq script's calligraphic and cursive features—where letter forms change based on where they appear in a word—make Urdu handwriting recognition a challenging process.  Urdu language has not been studied in this way except for a few languages such as English, Arabic, where some trends in handwriting recognition have been noted. This review is concerned with trends in deep learning abstraction, especially convolutional neural networks (CNNs) which have shown success in handwritten text recognition of Urdu. It has been highlighted though that there is a number of significant challenges such as the complex nature of the ligatures, diversity of the writing tendencies, as well as insufficient quantity of extensive annotated Urdi datasets. In spite of the above context, some research in the address recognition systems has taken advantage of an Urdu-Nastaleeq Handwritten Dataset (UNHD) and Urdu Handwritten Text Dataset (UHTD), as well as a Urdu Handwritten Character Database (UHCD) but the research works still remains unreliable since the datasets are limited. Such approaches enable the researcher to make some drawing similarities not only between Urdu language and Pashto language but also with printed Persian Language. It is evident in this review that the center of attention has been the recognition of Urdu handwritten text rather than all the main points including language and algorithms.