They're veGAN but they almost taste the same: generating simili-manuscripts with artificial intelligence
Fiche du document
10 juillet 2023
- handle: 10670/1.dprj8j
- hal: hal-04086940
http://creativecommons.org/licenses/by-sa/ , info:eu-repo/semantics/OpenAccess
Mots-clés
Computational humanities Computer vision Paleography Sciences humaines computationnelles Paléographie Vision par ordinateurSujets proches
CodicesCiter ce document
Chahan Vidal-Gorène et al., « They're veGAN but they almost taste the same: generating simili-manuscripts with artificial intelligence », HAL-SHS : linguistique, ID : 10670/1.dprj8j
Métriques
Partage / Export
Résumé
The aim of our research is to artificially generate fake historical manuscripts using generative adversarial networks, better known as GANs. At this stage, these experiments pursue two objectives: evaluating feasibility of generating realistic fake manuscripts under certain conditions of layout, script, or date, and creating artificial data for HTR training, as is done for printed materials with synthetic data. Ground truth creation is indeed a time consuming task, in particular for ancient languages for which we still lack specialists able to manually annotate documents. GAN appears as a relevant answer to this challenge, as they reach very convincing results in different scenarii. This short paper investigates the feasibility of the line-level approach for historical manuscripts.