Research On Classifier Combination And Its Applications To Human-Computer Interaction System

As the practical problems in the field of pattern recognition becomes more complex, the performance of one single classifier may not meet the demands in many real-world applications.Therefore,classifier combination has become a novel and important methodology to improve the performance of pattern recognition systems.Research on classifier combination methods is crucial for the development of human-computer interaction.Modern human-computer interaction technology has developed into the stage of multi-modal human-computer interaction,in which multi-modal recognition is an important topic.Above all,classifier combination is a key technology to solve multi-modal recognition problems.In this dissertation,the theories and algorithms of classifier combination are studied for the multi-modal recognition problems in human-computer interaction.The following four kinds of classifier combination problems are studied:(1) classifier combination problems with a large number of classes(the number of classes is larger than 15);(2) classifier combination problems with a small number of classes(the number of classes is between 3 and 15);(3) two-class classifier combination problems;(4) local-accuracy based classifier combination problems.This dissertation makes the following achievements:1.A novel kind of rank level classifier fusion method,named rank transformation method,is proposed for classifier combination problems with a large number of classes. By combining rank transformation with the weighted classifier fusion rule,the proposed method tries to increase the influence of the small rank values to the final classification results.Experimental results show that the proposed method outperforms traditional classifier combination methods by 0.1～1 percent in testing accuracy rate.2.A new measurement level classifier fusion method,namly multiple decision templates(MDT),is proposed for measurement level classifier fusion problems with a small number of classes.The proposed method can increase the classification accuracy efficiently by using each decision template to reduce one kind of classification error. ELENA data sets and UCI data sets are used to test the MDT method,and the experimental results show the good performences of the proposed method.In comparison with the existing classifier combination method,such as voting,naive Bayesian,decision templates,the testing accuracy rate of the proposed MDT method is about 0.4-0.9 percent higher.Moreover,compared with the k-NN rule,the proposed method demonstrates the comparable testing accuracy rate and the low computation amount when the number of training samles is sufficiently large.When the number of training samples decreases,the proposed method can achieve higher testing accuracy rate3.A new measurement level classifier fusion method named class-boundary based classifier fusion(CBCF) is proposed for two-class classifier combination problems. Based on the properties of the meta-level space of the considered problem,this method directly extracts the boundaries of the classes from the training set,and then defines local linear combination rules basing on the points on the boundaries.Experimental results on the Phoneme and Ringnorm data sets show that:compared with the existing classifier combination methods,such as voting,naive Bayesian,decision templates, and linear combination,the accuracy of the proposed CBCF method is 0.7～1.5 percent higher;compared with the k-NN rule,they are comparable in accuracy whereas the computational amount of the proposed method is 1/50-1/20 times than that of the k-NN rule.4.Towards the local-accuracy based classifier combination problems,a method transforming local accuracy to classification confidence is proposed.Therefore,one can utilize classifier fusion algorithm to replace the traditional dynamic classifier selection methods to achieve higher accuracy.After the new method implemented,dynamic classifier selection is equivalent to the Max rule,so that using more effective classifier fusion methods can achieve higher accuracy.Experimental results on the Elena,UCI and Ringnorm data sets show that the proposed method can improve the accuracy of the multi-classifier systems by 0.2～13.6 percent higher.5.A multi-modal person recognition system and a multi-modal person verification system are designed based on the new classifier combination methods.The multi-modal person recognition system uses a rank transformation method because it includes many classes.Meanwhile,the multi-modal person verification system includes only two classes,so the CBCF method is utilized.Experimental results on the multi-modal database show the better performances produced by the proposed classifier combination methods.The recognition rate of the person recognition system increases from 94%(the highest of the recognition rates of the base classifiers) to 99.71%,while the HTER(Half Total Error Rate) of the person verification system decreases from 5.12%(the least of the HTERs of the base classifiers) to 0.92%.Finally,a multi-modal person recognition system and a multi-modal person verification system are designed.The research work lays a sound basis for applications in multi-modal human-computer interaction.