Computer Engineering and Applications ›› 2019, Vol. 55 ›› Issue (11): 110-116.DOI: 10.3778/j.issn.1002-8331.1812-0121

Previous Articles     Next Articles

Establishment and Analysis of Large-Scale Continuous Chinese Sign Language Dataset

YUAN Tiantian,ZHAO Wei,YANG Xue,HU Bin   

  1. Technical College for the Deaf, Tianjin University of Technology, Tianjin 300384, China
  • Online:2019-06-01 Published:2019-05-30

大规模连续中国手语数据集的创建与分析

袁甜甜,赵  伟,杨  学,胡  彬   

  1. 天津理工大学 聋人工学院,天津 300384

Abstract: The vast majority of hearing people who do not understand sign language cause deaf people who find jobs, seek medical treatment, legal advice, or other life and work area great communication difficulties. And sign language interpreters need to be made appointments in advance, the cost is also very high. Therefore, many researchers have begun to use machine learning to develop automatic sign language translators, but most of the researches do not work well because of the size and quality of the data sets. To create the largest Chinese sign language data set in the world, the end-to-end learning model of body joints’ position, facial expression and finger joint is used to train effectively. Conclusion shows the great advantage of modern deep learning technology in the recognition of complex sign language is highlighted. The BLEU-4 for a small subset has reached 30.8.

Key words: sign language recognition, deep learning, dataset, feature extraction, end-to-end

摘要: 绝大多数健听人不懂手语导致听障人在找工作、就医、法律咨询等各生活、工作领域中遇到了极大的沟通障碍,而手语翻译员需要提前预约,成本也非常高,所以很多科研工作者都开始利用机器学习来开发手语自动翻译器,但其中的大部分研究都因为受到了数据集规模和质量的影响而效果不佳。为解决上述矛盾和问题,创建了目前全球最大的中国连续手语数据集,并使用了考虑身体关节的位置、面部表情及手指关节的端到端的深度学习模型进行有效训练。结论突显了现代深度学习技术在识别复杂手语方面的巨大优势,针对较小子集的BLEU-4已达到30.8。

关键词: 手语识别, 深度学习, 数据集, 特征提取, 端到端