UrduDeepNet: offline handwritten Urdu character recognition using deep neural network

Abstract

Handwritten Urdu character recognition system faces several challenges including the writer-dependent variations and non-availability of benchmark databases for cursive writing scripts. In this study, we propose a handwritten Urdu character dataset for Nasta�liq writing style covering isolated, positional characters as well as numerals. We also propose a convolutional neural network (CNN) architecture for the recognition of handwritten Urdu characters and numerals. CNN is a novel technique for image recognition that does not need explicit feature engineering and extraction and produces efficient results as compared to standard handcrafted feature extraction approaches. The proposed system was trained on a training dataset of 74, 285 samples and evaluated on a test dataset of 21, 223 samples and achieved a recognition rate of 98.82% for 133 classes, outperforming the results of all state-of-the-art systems for the Urdu language. � 2021, The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature.