TY - GEN
T1 - Thai Scene Text Recognition with Character Combination
AU - Li, Chun
AU - Zhan, Hongjian
AU - Zhao, Kun
AU - Lu, Yue
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022.
PY - 2022
Y1 - 2022
N2 - In recent years, scene text recognition(STR) that recognizing character sequences in natural images is in great demand beyond various fields. However, most STR studies only focus on popular scripts like Chinese or English, too little attention has been paid to minority languages. In this paper, we address problems on Thai STR, and introduce a novel strategy called Thai Character Combination(TCC), which explore original characteristics of Thai text. Unlike most other scripts, characters in Thai text can be written both horizontally and vertically, which brings big challenges to current sequence-based text recognition methods. In order to reduce complexity of structure and alleviate the misalignment problem in attention-based methods, TCC intends to combine Thai characters that stack vertically to independent combined characters. Furthermore, we establish a Thai Scene Text(TST) dataset that collected from multiple scenarios to evaluate the performance of our proposed character modeling strategy. We conduct abundant experiments and analyses to compare the recognition performance of models with and without TCC. The results indicate the effectiveness of the proposed method from multiple perspectives, especially, TCC benefits a lot for long text recognition, and there is a substantial improvement in the recognition accuracy of entire string-level.
AB - In recent years, scene text recognition(STR) that recognizing character sequences in natural images is in great demand beyond various fields. However, most STR studies only focus on popular scripts like Chinese or English, too little attention has been paid to minority languages. In this paper, we address problems on Thai STR, and introduce a novel strategy called Thai Character Combination(TCC), which explore original characteristics of Thai text. Unlike most other scripts, characters in Thai text can be written both horizontally and vertically, which brings big challenges to current sequence-based text recognition methods. In order to reduce complexity of structure and alleviate the misalignment problem in attention-based methods, TCC intends to combine Thai characters that stack vertically to independent combined characters. Furthermore, we establish a Thai Scene Text(TST) dataset that collected from multiple scenarios to evaluate the performance of our proposed character modeling strategy. We conduct abundant experiments and analyses to compare the recognition performance of models with and without TCC. The results indicate the effectiveness of the proposed method from multiple perspectives, especially, TCC benefits a lot for long text recognition, and there is a substantial improvement in the recognition accuracy of entire string-level.
KW - Scene text recognition
KW - Thai Character Combination
KW - Thai scene text dataset
UR - https://www.scopus.com/pages/publications/85142849380
U2 - 10.1007/978-3-031-18913-5_25
DO - 10.1007/978-3-031-18913-5_25
M3 - 会议稿件
AN - SCOPUS:85142849380
SN - 9783031189128
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 320
EP - 333
BT - Pattern Recognition and Computer Vision - 5th Chinese Conference, PRCV 2022, Proceedings
A2 - Yu, Shiqi
A2 - Zhang, Jianguo
A2 - Zhang, Zhaoxiang
A2 - Tan, Tieniu
A2 - Yuen, Pong C.
A2 - Guo, Yike
A2 - Han, Junwei
A2 - Lai, Jianhuang
PB - Springer Science and Business Media Deutschland GmbH
T2 - 5th Chinese Conference on Pattern Recognition and Computer Vision, PRCV 2022
Y2 - 4 November 2022 through 7 November 2022
ER -