非特定人手语识别统计模型的改进及应用

目前,非特定人手语识别与特定人系统相比还有较大的差距.手语数据差异性使得非特定人手语识别中提取手语数据有效的共同特征非常困难,因而,手语数据差异性在一定程度上影响了非特定人手语识别的识别效果.本文从手语数据存在差异性这一角度入手,利用流形概念的学习和推理能力并在流形允许变化的范围内进行有效建模.在建模的过程中,从范函求极值的角度出发,给出了一个让人容易理解且直观化的推导过程.进而应用流形概念中的切向量来改进手语识别的统计模型(TV/HMM)并应用于大词汇量非特定人手语识别,以解决手语数据的差异性对大词汇量非特定人手语识别所造成的影响.实验表明,改进后的TV/HMM识别系统在大词汇量非注册的易混词集上识别率高明显.

Improving signer-independent sign language recognition

There is a huge gap between signer-independent sign language recognition and signer-dependent sign language recognition systems. The data variance from different signers in sign language makes it difficult to extract effective common features of data in signer-independent recognition. This data variance unavoidably affects the effect of signer-independent sign language recognition. This paper presents a model of an allowable variance range. It uses the learning and reasoning abilities of manifold concept to deal with sign language data variance. An easy and intuitive derivation method was created to establish the extremum of the function. A manifold tangent vectors based sign language recognition statistical model (TV/HMM) was applied in our signer-independent sign language recognition to resolve data variance in signer-independent sign language recognition. Experiments showed that, compared with traditional HMM recognition systems, the average discrimination rate significantly improves.