JIHAD EL SANA

This paper presents Historical Style Transfer Generative Adversarial Networks (HST-GAN) for generating historical text images. Our model consists of three blocks: Encoder, Generator, and Discriminator. The Encoder encodes the style to be generated, and the Generator applies an encoded style, S to an input text image, I, and generates a new image with the content of I and the style S. The Discriminator encourages the Generator to enhance the quality of the generated images. Multiple loss functions are applied to ensure the generation of quality images. We evaluated our model against three challenging historical handwritten datasets of two different languages. In addition, we compare the performance of HST-GAN with the state of art approaches and show that HST-GAN provides the best generated images for the three tested datasets. We demonstrate the capability of HST-GAN to transfer multiple styles across domains by taking the style from one dataset and the content from another dataset and generate the content according to the desired style. We test the quality of the style domains transferring using a designated classifier and a human evaluation and show that the generated images are very similar to the original style.