A novel holistic unconstrained handwritten urdu recognition system using convolutional neural networks

Handwritten Urdu recognition has been the least explored to date due to unavailability of a standard hand-written Urdu dataset, huge variation among writing styles of different Urdu writers, irregular positioning of diacritics associated with ligatures, similarity in shape of some Urdu characters in writing, and unavailability of an efficient learning and training technique. Few researchers have proposed the handwritten Urdu datasets among which only Urdu Nastaliq handwritten dataset (UNHD) is publicly available. The UNHD contains ligatures of only up to five characters and does not cover the entire Urdu ligature corpus. Hence, we present a novel comprehensive handwritten Urdu dataset named UHLD for the ‘Urdu Handwritten Ligature Dataset’:—which consists of ligatures of up to seven-character length and covers most of the ligature corpus of the Urdu language. The UHLD is written by both genders independent of age of person, paper color, paper type (blank or ruled), ink color, pen type. We propose an unconstrained handwritten Urdu recognition system that can recognize handwritten Urdu ligatures with up to six characters. A new robust algorithm has also been proposed here that is able to divide a complete ligature into primary and secondary components with 98% accuracy on a large Urdu dataset. Our proposed holistic handwritten Urdu recognition system ensures independent recognition of both primary and secondary components of a word/ligature. The proposed recognition technique is transformation invariant and computationally efficient and achieves a better recognition rate of 97% for UHLD and 93% for UNHD.