Research On Pattern Classification Based On The D-S Evidence Theory

Pattern classification is a common problem in real life and it has developed to avery advanced stage. But at the same time, it has some problem to solve, such as thelimitations on too many characteristics of the classification model, the difficulties inobtaining samples, the difficult to determine by the obtained and uncertaininformation. Support vector machine (SVM) method which is based on statisticallearning theory is a new and effective methods in practical pattern classificationproblems. It has unique advantages and good prospects to solve the nonlinear, smallsample and high dimensional pattern recognition problem.D-S evidence theory which has a good performance in practical application is aset of mathematical method to handle uncertainty reasoning which is based on"evidence"and "combination" . It can solve the problem of uncertainty very well.Through the D-S evidence combination, make the uncertainty of the target which iswait for identification down, can improve the ability to identify targets effectively.Because of the obtained information with a considerable uncertainty in thepattern classification problems, so we can use the advantages to deal with uncertaininformation and data of D-S evidence theory and apply it to pattern classificationproblem. We can make one pattern classification problem to be completed by severalSVM classifiers and apply a reasonable combination to the output of those classifierswith D-S evidence theory. It can improve the classification accuracy, make the errorrecognition rate or loss smallest which is caused by classification of model which iswait for identification and solve the pattern classification to uncertain andhigh-dimensional data effectively.The research in this paper is to make the support vector machine combined withthe D-S evidence theory, exert respective advantages, get integrated model and solvethe classification problem better. The main context of the paper as follow:Expanding the standard of two-class SVM for multi-class SVM, and introducingthe posterior output probability of the SVM, we can get SVM with multi-class andprobability. And it has been proved that the classification performance of such SVMis much better than the hard decision method through experiments.Making use of the SVM with multi-class and probability to get basic probabilityassignment function, we can build support vector machines with many types ofevidence. In the process of fusion of evidence support vector machine by synthetic rules, the "one to one" strategy of evidence will appear evidence conflict in theintegration process. So this paper puts forward a kind of method to avoid conflictwhen the second use of Dempster rules.Applying the methods of output of non-probabilistic to define the basicprobability assignment function, this paper presents a model using real-valuedistance to compute weight factor of basic probability assignment function which isimproved on voting. Then using the D-S evidence theory fuses and gets the finalresult of judgment. Experimental results show that the proposed method is better thanthe original method.