Generative Adversarial Network for an Improved Arabic Handwritten Characters Recognition

Abstract Currently, Arabic character recognition remains one of the most complicated challenges in image processing and character identification. Many algorithms exist in neural networks, and one of the most interesting algorithms is called generative adversarial networks (GANs), where 2 neural networks fight against one another. A generative adversarial network has been successfully implemented in unsupervised learning and it led to outstanding achievements. Furthermore, this discriminator is used as a classifier in most generative adversarial networks by employing the binary sigmoid cross-entropy loss function. This research proposes employing sigmoid cross-entropy to recognize Arabic handwritten characters using multi-class GANs training algorithms. The proposed approach is evaluated on a dataset of 16800 Arabic handwritten characters. When compared to other approaches, the experimental results indicate that the multi-class GANs approach performed well in terms of recognizing Arabic handwritten characters as it is 99.7% accurate. Keywords: Generative Adversarial Networks (GANs), Arabic Characters, Optical Character Recognition, Convolutional Neural Networks (CNNs).