Publication detail

Finetuning Is a Surprisingly Effective Domain Adaptation Baseline in Handwriting Recognition

KOHÚT, J. HRADIŠ, M.

Original Title

Finetuning Is a Surprisingly Effective Domain Adaptation Baseline in Handwriting Recognition

Type

conference paper

Language

English

Original Abstract

In many machine learning tasks, a large general dataset and a small specialized dataset are available. In such situations, various domain adaptation methods can be used to adapt a general model to the target dataset. We show that in the case of neural networks trained for handwriting recognition using CTC, simple finetuning with data augmentation works surprisingly well in such scenarios and that it is resistant to overfitting even for very small target domain datasets. We evaluated the behavior of finetuning with respect to augmentation, training data size, and quality of the pre-trained network, both in writer-dependent and writer-independent settings. On a large real-world dataset, finetuning provided an average relative CER improvement of 25 % with 16 text lines for new writers and 50 % for 256 text lines.

Keywords

Handwritten text recognition, OCR, Data augmentation, Finetuning.

Authors

KOHÚT, J.; HRADIŠ, M.

Released

19. 8. 2023

Publisher

Springer Nature Switzerland AG

Location

San José

ISBN

978-3-031-41684-2

Book

Document Analysis and Recognition - ICDAR 2023

Edition

Lecture Notes in Computer Science

ISBN

0302-9743

Periodical

Lecture Notes in Computer Science

Year of study

14190

Number

1

State

Federal Republic of Germany

Pages from

269

Pages to

286

Pages count

18

URL

https://pero.fit.vutbr.cz/publications

BibTex

@inproceedings{BUT185151,
  author="Jan {Kohút} and Michal {Hradiš}",
  title="Finetuning Is a Surprisingly Effective Domain Adaptation Baseline in Handwriting Recognition",
  booktitle="Document Analysis and Recognition - ICDAR 2023",
  year="2023",
  series="Lecture Notes in Computer Science",
  journal="Lecture Notes in Computer Science",
  volume="14190",
  number="1",
  pages="269--286",
  publisher="Springer Nature Switzerland AG",
  address="San José",
  doi="10.1007/978-3-031-41685-9\{_}17",
  isbn="978-3-031-41684-2",
  issn="0302-9743",
  url="https://pero.fit.vutbr.cz/publications"
}