Detail publikace

TS-Net: OCR Trained to Switch Between Text Transcription Styles

KOHÚT, J. HRADIŠ, M.

Originální název

TS-Net: OCR Trained to Switch Between Text Transcription Styles

Typ

článek ve sborníku ve WoS nebo Scopus

Jazyk

angličtina

Originální abstrakt

Multiple transcribers produce transcriptions in inconsistent transcription styles.  This presents a problem for training consistent neural network systems for text recognition. We propose Transcription Style Block (TSB) which can learn to switch between multiple transcription styles without any explicit knowledge about the transcription rules. TSB is an adaptive instance normalization conditioned by transcription style identifiers e.g. document numbers or transcriber names and it can be added near the end of any standard text recognition network.  We show that TSB is robust towards the number and complexity of transcription styles and does not degrade the text recognition performance. With time and data efficient adaptation to a new transcription style, we achieved up to 77\% relative test character error reduction in comparison to a network without the TSB. 

Klíčová slova

Transcription styles, Adaptive instance normalization, Text recognition, Neural networks, CTC

Autoři

KOHÚT, J.; HRADIŠ, M.

Vydáno

9. 8. 2021

Nakladatel

Springer Nature Switzerland AG

Místo

Lausanne

ISBN

978-3-030-86336-4

Kniha

Lladós J., Lopresti D., Uchida S. (eds) Document Analysis and Recognition - ICDAR 2021

Edice

Lecture Notes in Computer Science

ISSN

0302-9743

Periodikum

Lecture Notes in Computer Science

Ročník

12824

Číslo

1

Stát

Spolková republika Německo

Strany od

478

Strany do

493

Strany počet

16

URL

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

BibTex

@inproceedings{BUT169806,
  author="Jan {Kohút} and Michal {Hradiš}",
  title="TS-Net: OCR Trained to Switch Between Text Transcription Styles",
  booktitle="Lladós J., Lopresti D., Uchida S. (eds) Document Analysis and Recognition - ICDAR 2021",
  year="2021",
  series="Lecture Notes in Computer Science",
  journal="Lecture Notes in Computer Science",
  volume="12824",
  number="1",
  pages="478--493",
  publisher="Springer Nature Switzerland AG",
  address="Lausanne",
  doi="10.1007/978-3-030-86337-1\{_}32",
  isbn="978-3-030-86336-4",
  issn="0302-9743",
  url="https://pero.fit.vutbr.cz/publications"
}