| On the research of pattern recogniton problems in the environment of a vast amount of data, it is difficult to build the corresponding classifiers, because these problems are usually summaried as follow: a feature space of high dimensionality for expressing these data and a data set with large amount of samples that belong to very many different classes and the decision boundaries in these problems are very complex. It is necessary not only to add different analysis and computing methods, but also to consider the integration among these methods .It is suggested that different classifiers offered complementary information,so we can combine classifiers to harness their strength. By means of LVQ and SVM as main tools, the integrated design for data of large scale is researched and discussed in the paper. The main contents of this paper are as follows: (1) Analysis the recent situations on application and developments about classifiers. The gneralization performance of the classifier is discussed. (2) Learning Vector Quantization(LVQ) algorithm has been widely used in the area of pattern recognition. Based on the anlysis of LVQ algorithm, the method of similar set is presented. This leads to the reduction of the training time of the LVQ algorithm.Then SLVQ is combined with the Fuzzy LVQ, which leads to a good recognition accuracy. (3) Based on the anlysis of the theory of SVM and methods of SVM multiclas, the DTSVM (decision tree based SVM) is designed to solve pattern recogniton problems in the environment of a vast amount of data, and the result is analyzed. (4) For the data of large scale, it is difficult to get ideal results with only a single classifier. So a multi-stage classifier is built. The multi-stage classifier is based on the improved LVQ and SVM. The method is to reduce the scale of problem in each stage and different classifiers offered complementary information. The result shows that this method can largely improve the recognition accuracy. |
PDF Full Text Request