Table of Content

Open Access

ARTICLE

CTSF: An End-to-End Efficient Neural Network for Chinese Text with Skeleton Feature

Hengyang Wang, Jin Liu*, Haoliang Ren
Shanghai Maritime University, Shanghai, 201306, China
* Corresponding Author: Jin Liu. Email:

Journal on Big Data 2021, 3(3), 119-126. https://doi.org/10.32604/jbd.2021.017184

Received 23 January 2021; Accepted 27 June 2021; Issue published 22 November 2021

Abstract

The past decade has seen the rapid development of text detection based on deep learning. However, current methods of Chinese character detection and recognition have proven to be poor. The accuracy of segmenting text boxes in natural scenes is not impressive. The reasons for this strait can be summarized into two points: the complexity of natural scenes and numerous types of Chinese characters. In response to these problems, we proposed a lightweight neural network architecture named CTSF. It consists of two modules, one is a text detection network that combines CTPN and the image feature extraction modules of PVANet, named CDSE. The other is a literacy network based on spatial pyramid pool and fusion of Chinese character skeleton features named SPPCNN-SF, so as to realize the text detection and recognition, respectively. Our model performs much better than the original model on ICDAR2011 and ICDAR2013 (achieved 85% and 88% F-measures) and enhanced the processing speed in training phase. In addition, our method achieves extremely performance on three Chinese datasets, with accuracy of 95.12%, 95.56% and 96.01%.

Keywords

Deep learning; convolutional neural network; Chinese character detection; text segmentation

Cite This Article

H. Wang, Jin Liu and H. Ren, "Ctsf: an end-to-end efficient neural network for chinese text with skeleton feature," Journal on Big Data, vol. 3, no.3, pp. 119–126, 2021.



This work is licensed under a Creative Commons Attribution 4.0 International License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
  • 583

    View

  • 419

    Download

  • 0

    Like

Share Link

WeChat scan