Paper Abstract and Keywords |
Presentation |
2023-03-02 17:00
A Semi-Supervised Learning Framework for Handwritten Text Recognition using Mixed Augmentations and Scheduled Pseudo-Label Loss Masayuki Honda, Hung Tuan Nguyen, Cuong Tuan Nguyen (TUAT), Cong Kha Nguyen, Ryosuke Odate, Takashi Kanemaru (Hitachi Ltd.), Masaki Nakagawa (TUAT) PRMU2022-97 IBISML2022-104 |
Abstract |
(in Japanese) |
(See Japanese page) |
(in English) |
We propose Incremental Teacher Model, a semi-supervised learning (SSL) framework for handwriting text recognition. The framework comprises a teacher model and a student model learning through Scheduled Pseudo-Label loss with Mixed Augmentations. First, the student model is pre-trained by labeled samples and used to initiate the teacher model. The student model is further trained on transformed samples provided by Mixed Augmentations. This training process uses pseudo-labels generated by the teacher model. After a training epoch, the teacher model is updated from a well-validated student model. We apply the proposed framework to four architectures of different handwriting recognizers. For almost every architecture, the recognizer trained by Incremental Teacher Model outperforms the recognizers trained by the well-known SSL methods on the IAM handwriting database. |
Keyword |
(in Japanese) |
(See Japanese page) |
(in English) |
Semi-Supervised Learning / Training framework / Handwriting recognition / / / / / |
Reference Info. |
IEICE Tech. Rep., vol. 122, no. 404, PRMU2022-97, pp. 199-204, March 2023. |
Paper # |
PRMU2022-97 |
Date of Issue |
2023-02-23 (PRMU, IBISML) |
ISSN |
Online edition: ISSN 2432-6380 |
Copyright and reproduction |
All rights are reserved and no part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the publisher. Notwithstanding, instructors are permitted to photocopy isolated articles for noncommercial classroom use without fee. (License No.: 10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
Download PDF |
PRMU2022-97 IBISML2022-104 |
|