TY - GEN
T1 - Text Enhancement for Historical Handwritten Documents
AU - Alaasam, Reem
AU - Madi, Boraq
AU - El-Sana, Jihad
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024.
PY - 2024/1/1
Y1 - 2024/1/1
N2 - This paper presents a text enhancement method for historical handwritten documents. Text enhancement is a sub-field of super-resolution focused on document images and text. Our work is based on generative adversarial networks (GAN), and we have introduced a new evaluation metric for document image enhancement. Our approach denoises historical documents and holistically increases their resolution using GANs. We modified The generator structure of our GAN model by replacing Batch Norm layers with Residual-in-Residual Dense Blocks (RRDB) and adopting a discriminator based on the Relativistic GAN. Our evaluation metric for text enhancement focuses on text quality based on the magnitude of gradients at text edges to assess the improvement of the generated images. We tested our method on three degraded handwritten historical datasets of two languages and obtained excellent results. In addition, We compare our approach with SRGAN, Nearest, and Bi-Cubic interpolations and show that our method performs much better than these methods on the three datasets. The proposed approach can handle various types of noise while applying text enhancement up to 16 times the input image.
AB - This paper presents a text enhancement method for historical handwritten documents. Text enhancement is a sub-field of super-resolution focused on document images and text. Our work is based on generative adversarial networks (GAN), and we have introduced a new evaluation metric for document image enhancement. Our approach denoises historical documents and holistically increases their resolution using GANs. We modified The generator structure of our GAN model by replacing Batch Norm layers with Residual-in-Residual Dense Blocks (RRDB) and adopting a discriminator based on the Relativistic GAN. Our evaluation metric for text enhancement focuses on text quality based on the magnitude of gradients at text edges to assess the improvement of the generated images. We tested our method on three degraded handwritten historical datasets of two languages and obtained excellent results. In addition, We compare our approach with SRGAN, Nearest, and Bi-Cubic interpolations and show that our method performs much better than these methods on the three datasets. The proposed approach can handle various types of noise while applying text enhancement up to 16 times the input image.
KW - Generative Adversarial Networks
KW - Historical Handwritten Documents
KW - Super-Resolution
UR - http://www.scopus.com/inward/record.url?scp=85204379337&partnerID=8YFLogxK
U2 - 10.1007/978-3-031-70536-6_24
DO - 10.1007/978-3-031-70536-6_24
M3 - Conference contribution
AN - SCOPUS:85204379337
SN - 9783031705359
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 397
EP - 412
BT - Document Analysis and Recognition - ICDAR 2024 - 18th International Conference, Proceedings
A2 - Barney Smith, Elisa H.
A2 - Liwicki, Marcus
A2 - Peng, Liangrui
PB - Springer Science and Business Media Deutschland GmbH
T2 - 18th International Conference on Document Analysis and Recognition, ICDAR 2024
Y2 - 30 August 2024 through 4 September 2024
ER -