Text Enhancement for Historical Handwritten Documents

Reem Alaasam, Boraq Madi, Jihad El-Sana

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

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 Our approach denoises historical documents and holistically increases their resolution using GANs.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.

Original languageEnglish
Title of host publicationDocument Analysis and Recognition - ICDAR 2024 - 18th International Conference, Proceedings
EditorsElisa H. Barney Smith, Marcus Liwicki, Liangrui Peng
PublisherSpringer Science and Business Media Deutschland GmbH
Pages397-412
Number of pages16
ISBN (Print)9783031705359
DOIs
StatePublished - 1 Jan 2024
Event18th International Conference on Document Analysis and Recognition, ICDAR 2024 - Athens, Greece
Duration: 30 Aug 20244 Sep 2024

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume14805 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference18th International Conference on Document Analysis and Recognition, ICDAR 2024
Country/TerritoryGreece
CityAthens
Period30/08/244/09/24

Keywords

  • Generative Adversarial Networks
  • Historical Handwritten Documents
  • Super-Resolution

ASJC Scopus subject areas

  • Theoretical Computer Science
  • General Computer Science

Fingerprint

Dive into the research topics of 'Text Enhancement for Historical Handwritten Documents'. Together they form a unique fingerprint.

Cite this