Arabic Handwritten Characters Recognition using Convolutional Neural Network

Abstract

Recognition of Arabic handwritten characters is very important due to its various benefits and usages. Ancient documents, bank processing, postal mailing and others are examples where we may need character recognition systems. But many obstacles may be faced due to diversity of human writing styles. Language characters recognition has been widely covered in many languages and many algorithms and paradigms were used. With the strong appealing of deep CNN classifier promise results were reached in many classification problems. CNN is a feed forward neural network that is extensively used in several applications such as image classification. The main benefit of using CNN is the merging of feature extraction and classification itself. Some researchers used CNN in Arabic character recognition, one of those El-Sawy et al [1] who applied CNN architecture on a dataset namely (AHCD) of 16800 characters. They obtained a good accuracy of 94.9% and a misclassification error of 5.1% on testing data. In our paper we will explore their dataset by proposing a modified CNN architecture hopefully to overcome their results.