节点文献
面向大词汇量的连续中国手语识别系统的研究与实现
【作者】 王春立；
【导师】 高文；
【作者基本信息】 大连理工大学 ， 计算机应用技术， 2003， 博士
【摘要】 手语识别研究的目的是增进聋人与听力健康人之间无障碍的交流，提高计算机对人体语言的理解能力。由于手语使用范围不广，因此关于手语的研究较少，手语识别的研究也就只有10多年的历史。在手语识别领域中仍然有很多挑战性的难题，其中包括：大词汇量的手势语识别方法；词汇集可扩展的手语识别方法；手语识别的最小识别基元；过渡帧的有效处理方法；非特定人的手语识别方法。以上5个问题的解决对手语识别具有非常重要的意义。本文对以上5个问题中的前4个进行了研究，并在此基础上实现了大词汇量手语识别系统。 首先，文中对多数据流CHMM、多数据流SCHMM、多数据流DHMM和基于流捆绑技术的HMM这四种方法进行了对比，选择出最后一种方法作为大词汇量手语识别的核心技术。 其次，本文采用基于流状态捆绑的手语识别方法，并利用了基于动态规划算法的自动估计状态结点数的模型、修正转移矩阵、快速匹配、估计跳转参数等技术来提高系统的性能，在本文采集的世界上最大的手势库上(5100个手势)进行测试，实验结果表明这种方法是十分有效的。 此外，通过对手语辞典的分析，本文归纳整理出2400多个词根，并实现了基于词根的识别系统，在这个系统中，使用了树状搜索结构、前向索引表和N-Best方法。并将这种方法与基于手势词的方法进行了比较。最后，还对进一步寻找更基本的识别单元做了尝试。
【Abstract】 The aim of research on sign language recognition is to enable the communication between the hearing impaired and hearing-enabled to be free and to enhance the capacity of body language understanding of computers. As the sign language is not used widely, little research on it is held, and the research history of sign language recognition is only about ten years. There are many challenges in sign language recognition, which include: How to realize the recognition of large vocabulary sign language? What is the sign language recognition approach with scalability about the vocabularies? What are the smallest recognition units? How to solve the movement between two signs? How to solve problems for the signer independent sign language recognition? The solutions of the five problems mentioned above are very important to sign language recognition. Only four preceding problems are investigated in this paper. Based on this, a signer dependent large vocabulary sign language recognition system was designed and implemented.Firstly, the comparisons between multi-stream CHMM, multi-stream SCHMM, multi-DHMM and HMM based on banding technique are done here, and the last one is chosen as the kernel technique of the large vocabulary sign language recognition system.Secondly, the siga language recognition approach with scalability about the vocabularies is proposed based on the stream state tying. In order to improve the performance of the system, some useful ideas are employed, including building HMMs which can automatically estimate the numbers of the states based on dynamic programming, modifying the transferring probability, fast matching and using the estimating parameter of transferring between signs in search algorithm. The approach was test on the largest scale hand gesture database (5100 hand gestures) in the world. The experimental results have shown that this approach is very efficient.Thirdly, through analysis, about 2400 subwords are defined for CSL. A continuous sign language recognition approach based on subwords is proposed. Tree-structured network, the forward index table and N-Best are used. The method is compared to that based on the sign. At last, some methods to find the smaller subwords are studied.