Recognition of Urdu Handwritten Words Using Deep Learning Techniques

Cursive hand script recognition is reported as a problematic task in words recognition due to variations in writing styles from person to person. Urdu script has related challenges comparable to Arabic script due to the complex ligatures. To cope with these issues, we proposed to use deep learning (DL) techniques i.e., transfer learning and ensemble learning. A novel dataset was generated which comprised 2496 images of Urdu handwritten words by 26 different writers and 288 images of computer-generated words. The whole dataset contained 96 classes. The data augmentation technique was used to add diversity and increase the dataset size to 13,920 images. To thoroughly evaluate the results of this study, a comparative analysis of the accomplishments was carried out. Among all the applied DL models, the ensemble learning model produced the best results as compared to state-of-the-art models. The proposed model moves handwritten words recognition to a higher level and can efficiently be automated.