非特定人手语识别参数训练模型的改进及应用  

Improving and applying a signer-independent sign language recognition parameter training model

作　　者：倪训博[1] 程丹松[1] 吕海峰 王克家 耿铁珍[1] 

机构地区：[1]哈尔滨工业大学计算机学院,黑龙江哈尔滨150001 [2]中广核风力发电有限公司,北京100070 [3]哈尔滨工程大学自动化学院,黑龙江哈尔滨150001

出　　处：《哈尔滨工程大学学报》2009年第9期1035-1040,共6页Journal of Harbin Engineering University

基　　金：国家自然科学基金重点资助项目(60533030);国家自然科学基金资助项目(60873142,60602007);哈尔滨市科技创新人才基金资助项目(2008rfqxs037)

摘　　要：在手语识别研究中,非特定人手语识别参数训练的样本缺乏影响了非特定人手语识别的识别率.区分性训练可以很好地弥补由于训练样本的缺乏对识别系统所造成的影响,能够提高非特定人手语识别的识别率.对区分性训练(DT)所改进的HMM参数训练模型(DT/HMM)做了全新的推导,获得了与HMM相一致齐全的DT/HMM的参数模型.在特定人识别系统上应用可区分性训练的h准则获取了h参数,将该齐全的DT/HMM参数训练模型和h参数应用于大词汇量的非特定人手语识别当中,加入主观经验后的非注册易混词集EXP与MLE和EBW的非注册易混词集相比,平均识别率分别提高了10.65%和9.55%.In sign language recognition, a lack of training samples for signer-independent sign language decreases recognition rates due to an inability to identify suitable parameters. Discriminative training methods can improve the impact of insufficient training samples on the recognition system while increasing the recognition rate of signer-independent sign language recognition （SISLR）. Hidden Markov model （HMM） and dependency-tree hidden Markov model （DT-HMM） improvements through discriminative training were proven theoretically possible, so a DT-HMM model with complete parameters was derived and proven to be consistent with the HMM model. We obtained the h parameter by applying the h criterion of discriminative training to recognition systems optimized for specific people. A full range of DT-HMM parameter model consistent with HMM has been deduced in this paper. The h parameters are worked out by applying the h rcriterion of discriminative training method to a signer-dependent sign language recognition. Then, applying the full range of DT-HMM parameter model in a large vocabulary of words for signer- independent sign language recognition（SISLR） , to EXP, the average rates of recognition increase 10.65% and 9.55% compare with the nonregistered confusable set of MLE and EBW respectively.

关 键 词：区分性训练 DT/HMM 非特定人手语识别 参数训练模型 h参数 h准则 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]