非特定人手语识别参数训练模型的改进及应用

摘要

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 signerindependent 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.%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的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的参数训练模型和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%.