Deep Learning Application for Handwritten Arabic Word Recognition

Abstract

Automatic handwriting recognition is the process of converting online and offline letters or words as a graphical form into its text format. Automatic Arabic handwriting words recognition using deep learning neural networks is still in the early stages in terms of research. There are no general, complete, and reliable Arabic Handwritten words database (lexicon) that can be used as a reference or a benchmark for all researchers who want to extend the work on automatic Arabic handwriting word recognition. Also, many historic Arabic manuscripts have deteriorated because of inappropriate storage and most of them have not been digitized due to the lack of reliable database that can be used to recognize the words of Arabic manuscripts. Deep Convolutional Neural Networks (DCNNs) can be used to solve the problems of automatic Arabic handwriting words recognition. In this work, a new DCNN algorithm applied to a new dataset of Handwritten Arabic words representing the seven days of the week named Arabic Handwritten Weekdays Dataset (AHWD) has been programmed, tested, and analyzed. Our dataset contains 21357 words equally distributed between the seven classes and prepared by 1000. So, it can be used for training and testing on a reliable DCNN model that will be able, after training to generalize to new datasets. The model works by training a (DCNN) model on a balanced-randomly-selected dataset using different structures. The results are improved by adding drop-out, image regularization, proper learning rate to avoid overfitting of the data. Finally, a blind test has been performed on the hidden test set and the performance was reported using a confusion matrix and learning curves as a validation tool for the model. Results show that our model’s performance is promising, achieving accuracy rate of 0.9939 with error rate of 0.0461 using AHWD dataset, and accuracy rate of 0.9971 with error rate of 0.0171 using IFN/ENIT dataset.