Training transformer architectures on few annotated data: an application to historical handwritten text recognition - Nantes Université Access content directly
Journal Articles International Journal on Document Analysis and Recognition Year : 2024

Training transformer architectures on few annotated data: an application to historical handwritten text recognition

Abstract

Transformer-based architectures show excellent results on the task of handwritten text recognition, becoming the standard architecture for modern datasets. However, they require a significant amount of annotated data to achieve competitive results. They typically rely on synthetic data to solve this problem. Historical handwritten text recognition represents a challenging task due to degradations, specific handwritings for which few examples are available and ancient languages that vary over time. These limitations also make it difficult to generate realistic synthetic data. Given sufficient and appropriate data, Transformer-based architectures could alleviate these concerns, thanks to their ability to have a global view of textual images and their language modeling capabilities. In this paper, we propose the use of a lightweight Transformer model to tackle the task of historical handwritten text recognition. To train the architecture, we introduce realistic looking synthetic data reproducing the style of historical handwritings. We present a specific strategy, both for training and prediction, to deal with historical documents, where only a limited amount of training data is available. We evaluate our approach on the ICFHR 2018 READ dataset which is dedicated to handwriting recognition in specific historical documents. The results show that our Transformer-based approach is able to outperform existing methods.
Embargoed file
Embargoed file
0 5 1
Year Month Jours
Avant la publication
Thursday, July 25, 2024
Embargoed file
Thursday, July 25, 2024
Please log in to request access to the document

Dates and versions

hal-04447488 , version 1 (09-02-2024)

Identifiers

Cite

Killian Barrere, Yann Soullard, Aurélie Lemaitre, Bertrand Coüasnon. Training transformer architectures on few annotated data: an application to historical handwritten text recognition. International Journal on Document Analysis and Recognition, inPress, International Journal on Document Analysis and Recognition (IJDAR), ⟨10.1007/s10032-023-00459-2⟩. ⟨hal-04447488⟩
0 View
0 Download

Altmetric

Share

Gmail Facebook X LinkedIn More