TY - GEN
T1 - HST-GAN
T2 - 15th IAPR International Workshop on Document Analysis Systems, DAS 2022
AU - Madi, Boraq
AU - Alaasam, Reem
AU - Droby, Ahmad
AU - El-Sana, Jihad
N1 - Publisher Copyright:
© 2022, Springer Nature Switzerland AG.
PY - 2022/1/1
Y1 - 2022/1/1
N2 - 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.
AB - 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.
KW - Generating document images
KW - Generative adversarial networks
KW - Historical handwritten styles transfer
UR - http://www.scopus.com/inward/record.url?scp=85131114934&partnerID=8YFLogxK
U2 - 10.1007/978-3-031-06555-2_35
DO - 10.1007/978-3-031-06555-2_35
M3 - Conference contribution
AN - SCOPUS:85131114934
SN - 9783031065545
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 523
EP - 537
BT - Document Analysis Systems - 15th IAPR International Workshop, DAS 2022, Proceedings
A2 - Uchida, Seiichi
A2 - Barney, Elisa
A2 - Eglin, Véronique
PB - Springer Science and Business Media Deutschland GmbH
Y2 - 22 May 2022 through 25 May 2022
ER -