Overview

Original Research

ARABIC HANDWRITTEN CHARACTER CLASSIFICATION AND RECOGNITION USING CNN AND TRANSFER LEARNING APPROACH

OMAR ALI BORAIK 1, M. RAVIKUMAR 2, and CHANNABASAVA CHOLA 3.

Vol 17, No 11 ( 2022 )   |  DOI: 10.5281/zenodo.7393430   |   Author Affiliation: Department of Computer Science, Seiyun University, Hadramout, Yemen, Department of Computer Science and MCA, Kuvempu University, Shankaraghatta, Shimoga, Karnataka, India 1; Department of Computer Science and MCA, Kuvempu University, Shankaraghatta, Shimoga, Karnataka, India 2; Department of Computer Science and Engineering, Indian Institute of Information Technology, Kottayam, India 3.   |   Licensing: CC 4.0   |   Pg no: 2087-2107   |   To cite: OMAR ALI BORAIK, et al., (2022). ARABIC HANDWRITTEN CHARACTER CLASSIFICATION AND RECOGNITION USING CNN AND TRANSFER LEARNING APPROACH. 17(11), 2087–2107. https://doi.org/10.5281/zenodo.7393430   |   Published on: 30-11-2022

Abstract

Arabic Handwritten character recognition is still a complex task. Lack of Arabic handwritten database is considered a huge problem in improving a new model for Arabic character recognition task or even proposing a new one. The remarkable similarity between Arabic letters' shapes, especially the handwritten scripts lead to the misclassification of Arabic characters. Though, there are a lot of efforts that have been made in Object Character Recognition (OCR) systems for Arabic text. Recently, these efforts have only reached partial use of classification and recognition. Progress in recognizing Arabic handwritten scripts is still unsatisfactory. Newly, deep learning methods are utilized, which make significant progress in identifying Arabic texts. This proposed study, deep learning-based models namely the ResNet50, DenseNet 121, VGG16 and CNN models were used. Also, a transfer-learning approach was applied for classifying and recognizing Arabic handwritten characters. The current study aims to get benefit from the pre-trained transfer-learning approaches on the ImageNet dataset. Then, it re-trains the approaches on a collected dataset, which includes 131,000 images of Arabic handwritten characters. The results of the models of character recognition with Inception V3 and ResneNet50 are improved. So, it is easy to generalize to similar problems of classifying handwritten characters. Therefore, the proposed models on the collected dataset in terms of overall accuracy, recall, precision, and F1-score by 98.94%, 100%, 93.75%, and 96.43%, respectively. Overall classification accuracy of 98.95% is achieved. Therefore, the proposed models on the collected dataset in terms of overall accuracy, recall, precision, and F1-score by 98.94%, 100%, 93.75%, and 96.43%, respectively. Permanently, the overall classification accuracy of 98.95% is achieved.

Keywords

Arabic character Recognition; Deep learning; Character classification; Transfer learning; CNN.