第4期 杨国亮等：基于改进MMⅫ的HMM训练算法及其在面部表情识别中的应用 .437. 反向训练，这是与Baum Welch一个显著区别. [5]茅晓泉，胡光锐.基于最大互信息的离散隐马尔可夫模型训练 方法，上海交通大学学报，2001,35(11)：1713 参考文献 [6]Juang B H.Chou W.Lee C H.Minimum classification error [1]Baum L E.Eagon JA.An inequality with applications to statisti- methods for speech recognition.IEEE Trans Speech Audio Pro- ces,1997,5(3):257 cal estimation for probabilistic functions of Markov processes and to a model for ecology.Bull Am Math Stat.1967.37(2):360 [7]Bahl L R.Brown P F,Souza P V,et al.A new algorithm for the [2]Baggenstoss P M.A modified Baum-Welch algorithm for hidden estimation of hidden Markoy model parametersIEEE Int. Markov models with multiple observation speeches.Int J Speech Conf.on Acousties.Speech,and Signal Processing.New York. 1998.493 Commun,2000,5:411 [3]Bahl L R,Brown P F,Mercer R L.Maximum mutual informa- [8]He Q H.Kwong S.A adaptation of hidden Markov models using tion estimation of hidden Markov model parameters for speech maximum model distance algorithm.IEEE Trans System Man recognition//IEEE Int.Conf.on Acoustics.Speech.and Signal Cybern,2004,34(2):270 Processing.Tokyo,1986:49 [9]Lucas B.Kanade T.An iterative image registration technique [4]Assaf B Y,Brushtein D.A discriminative training algorithm for with an application to stereo vision//Proc.of the 7th Internation- hidden Markov models.IEEE Trans Speech Audio Process, al Joint Conference on Artificial Intelligence.Vancouver,1981: 121 2004,12(3):204 HMM training algorithm based improved MMI and its application in facial expres- sion recognition YANG Guoliang2),WANG Zhiliang?),LIU Jiwei),WANG Guojiang?),CHEN Fengjun2) 1)Mechanical and Electrical Engineering School.Jiangxi University of Science and Technology,Ganzhou 341000,China 2)Information Engineering School.University of Science Technology Beijing Beijing 100083.China ABSTRACI A new approach for hidden Markov model(HMM)training based on an improved maximum mu- tual information (MMI)criterion was presented and HMM parameter adjustment rules were induced.By adopt- ing a more realistic MMI definition,discriminative information contained in the training data could be used to improve the performance of HM M and this method was also used in facial expression recognition.Facial expres- sion feature vector flows were extracted by using the improved optical flow algorithm,and a hybrid classifier based on the improved HMM and BP neural network was designed.Experimental results show that the new method provides satisfactory recognition performance and the method is powerful for HM M parameter estima- tion. KEY WORDS maximum mutual information criterion;hidden Markov model;optical flow algorithm;facial expression recognition反向训练‚这是与 Baum—Welch 一个显著区别． 参 考 文 献 ［1］ Baum L E‚Eagon J A．An inequality with applications to statisti￾cal estimation for probabilistic functions of Markov processes and to a model for ecology．Bull Am Math Stat‚1967‚37（2）：360 ［2］ Baggenstoss P M．A modified Baum-Welch algorithm for hidden Markov models with multiple observation speeches．Int J Speech Commun‚2000‚5：411 ［3］ Bahl L R‚Brown P F‚Mercer R L．Maximum mutual informa￾tion estimation of hidden Markov model parameters for speech recognition∥IEEE Int．Conf．on Acoustics‚Speech‚and Signal Processing．Tokyo‚1986：49 ［4］ Assaf B Y‚Brushtein D．A discriminative training algorithm for hidden Markov models．IEEE Trans Speech Audio Process‚ 2004‚12（3）：204 ［5］ 茅晓泉‚胡光锐．基于最大互信息的离散隐马尔可夫模型训练 方法．上海交通大学学报‚2001‚35（11）：1713 ［6］ Juang B H‚Chou W‚Lee C H．Minimum classification error methods for speech recognition．IEEE Trans Speech Audio Pro￾cess‚1997‚5（3）：257 ［7］ Bahl L R‚Brown P F‚Souza P V‚et al．A new algorithm for the estimation of hidden Markov model parameters ∥ IEEE Int． Conf．on Acoustics‚Speech‚and Signal Processing．New York‚ 1998：493 ［8］ He Q H‚Kwong S．A adaptation of hidden Markov models using maximum model distance algorithm．IEEE Trans System Man Cybern‚2004‚34（2）：270 ［9］ Lucas B‚Kanade T．An iterative image registration technique with an application to stereo vision∥Proc．of the7th Internation￾al Joint Conference on Artificial Intelligence．Vancouver‚1981： 121 HMM training algorithm based improved MMI and its application in facial expres￾sion recognition Y A NG Guoliang 1‚2）‚WA NG Zhiliang 2）‚LIU Jiwei 2）‚WA NG Guojiang 2）‚CHEN Fengjun 2） 1） Mechanical and Electrical Engineering School‚Jiangxi University of Science and Technology‚Ganzhou341000‚China 2） Information Engineering School‚University of Science ＆ Technology Beijing‚Beijing100083‚China ABSTRACT A new approach for hidden Markov model （HMM） training based on an improved maximum mu￾tual information （MMI） criterion was presented and HMM parameter adjustment rules were induced．By adopt￾ing a more realistic MMI definition‚discriminative information contained in the training data could be used to improve the performance of HMM and this method was also used in facial expression recognition．Facial expres￾sion feature vector flows were extracted by using the improved optical flow algorithm‚and a hybrid classifier based on the improved HMM and BP neural network was designed．Experimental results show that the new method provides satisfactory recognition performance and the method is powerful for HMM parameter estima￾tion． KEY WORDS maximum mutual information criterion；hidden Markov model；optical flow algorithm；facial expression recognition 第4期 杨国亮等： 基于改进 MMI 的 HMM 训练算法及其在面部表情识别中的应用 ·437·