Arabic handwritten recognition based on deep convolutional neural network
Ala' Alzebdeh, M. Moneb Khaled, Mohammed Lataifeh, Ashraf Elnagar · IET conference proceedings. · 2022
Automatic handwritten recognition is a simulation of the human reading process by converting a handwritten input to an editable typed form. A successful deployment of such a system requires a robust character handwritten recognition module. Research papers conducted on the Arabic language are scarce compared to other languages due to the complexity and cursive nature of the Arabic language where characters are written in a conjoined and owing manner. We propose two deep learning models based on a Convolutional Neural Network (CNN) to classify Arabic handwritten characters using two publicly available datasets AIA9K and AHCD. Pre-processing and data augmentation have been successfully utilized to improve performance. We employed an Arabic word segmentation algorithm to disassemble a word into its letters, which makes the input to the classification system. The experimental results demonstrate solid performance with an average accuracy of 96.72% and 98.21% on AIA9k and AHCD datasets, respectively. The performance is boosted to 97.54% and 98.82%, with data augmentation on the respective datasets. Furthermore, the two augmented datasets were merged to build a generalized model that achieved outstanding accuracy of 98.27% using a modified approach of LeNet-5.