Conditional Deep Convolutional Generative Adversarial Networks for Isolated Handwritten Arabic Character Generation

Being the basis on which several languages of the world are built, the historical relevance of the basic Arabic characters cannot be overemphasized. Unique in its many similar characters which are only distinguishable by dots, Arabic character recognition and classification has witnessed notable increase in research in recent times, particularly using machine learning-based approaches. However, little or no research exists on automatic generation of handwritten Arabic characters. Besides, the available databases of labeled handwritten Arabic characters are limited. Motivated by this open area of research, we propose a Conditional Deep Convolutional Generative Adversarial Networks (CDCGAN) for a guided generation of isolated handwritten Arabic characters. Experimental findings based on qualitative and quantitative results show that CDCGAN produce synthetic handwritten Arabic characters that are comparable to the ground truth, given a mean multiscale structural similarity (MS-SSIM) score of 0.635 as against 0.614 in the real samples. Comparison with handwritten English alphabets generation task further shows the capability of CDCGAN in generating diverse yet high-quality images of handwritten Arabic characters despite their inherent complexity. Additionally, machine learning efficacy test using CDCGAN-generated samples shows impressive performance with about 10% performance gap between real and generated handwritten Arabic characters.