Generating Arabic Handwritten Characters Using DGAN

The aim of this experiment is to use deep convolutional generative adversarial network, which called (DGAN) to generate Arabic hand written character by changing the dataset of the original system (in google colab) from MNIST to the 12,440 Arabic handwritten character dataset collected by El-Sawy, Loey and El-Bakry in 2017. The main idea is to investigate the effectiveness of using DGN to generate Arabic handwritten characters. The image dimension of the Arabic handwritten dataset is 32x32x1, which is larger than the MNIST dataset images. Moreover, the Arabic handwritten dataset has 28 character classes, providing more variation than the MNIST handwritten dataset digits. The Keras sequential API is utilized to train the Generator and Discriminator using the Arabic handwritten dataset.

To manage the checkpoints, TensorFlow’s checkpointManager was used to restore the latest checkpoint before each training session. The checkpoint files and output images were stored in my Google Drive which was beneficial for running this Notebook on Google’s Colab across multiple sessions.
Using the same GAN as the tutorial with expanding the training EPOCHS until the generated images start looking like the real Image. The output in after 500 epochs looks acceptable in
terms of generating handwritten characters, and some of them look like Arabic characters as my observation.

References:
El-Sawy, A., Loey, M. and El-Bakry, H. (2017) Arabic Handwritten Characters Recognition using Convolutional Neural Network.