Urdu Handwritten Character Recognition using Deep Learning

  • Ali Jan Department of Computer Science & IT, University of Balochistan, Quetta, Pakistan
  • Junaid Baber Department of Computer Science & IT, University of Balochistan, Quetta, Pakistan
  • Anwar Ali Sanjrani Department of Computer Science & IT, University of Balochistan, Quetta, Pakistan
  • Maheen Bakhtyar Department of Computer Science & IT, University of Balochistan, Quetta, Pakistan
Keywords: Deep Learning, Urdu Language Processing, Optical Character Recognition

Abstract

Due to the emergence of ICT technologies, teaching and learning through the smart system is very trending. Kids start learning the basic alphabets of any language through the smart interaction of tablets and phones. However, very little research is reported on the Urdu language for such applications. Urdu is the national language of Pakistan and spoken across the subcontinent. In this paper, two famous deep learning-based models are evaluated for handwritten Urdu alphabets. Handwritten data is generated by the students and the teachers. Customized AlexNet is trained on these data, and accuracy is compared with the FERDNN model. The AlexNet is a famous model for various applications, and FERDNN is proposed for facial expressions. It has been observed during experiments that the AlexNet provides low accuracy using the default configurations in the first layer. Experiments show that AlexNet achieves competitive performance by changing the kernel size on the first layer, the customized AlexNet and FERDNN provides 0.95 and 0.83 accuracies on the training set, and 0.52 and 0.61 accuracies on the test set, respectively.

Published
2020-04-20
How to Cite
[1]
A. Jan, J. Baber, A. A. Sanjrani, and M. Bakhtyar, “Urdu Handwritten Character Recognition using Deep Learning”, jictra, pp. 42-51, Apr. 2020.
Section
Original Articles