Design And Combination On Classifier

Pattern recognition is an important aspect in the domain of artificial intelligence. Pattern Recognition system is composed of two parts: one is how to extract feature from pattern space, and transfer patterns to feature. The other is classification which according to the selected feature, and assign the input pattern to a certain class. We do some researches on the second part of pattern recognition about how to improve recognition performance.In the thesis, the method of minimum distance classifier is analyzed and researched, and its disadvantage is pointed out. In the minimum distance classifier, through analyzing Euclidean distance metric and introducing adaptive distance metric, a new training method based on adaptive distance metric is proposed. First, the method builds a model about adaptive distance metric using training samples. The model assures that the distance between training samples and the same pattern classification is the nearest, and the distance between training samples and other pattern classification is the farthest, then an optimal weight is obtained through solving objective function. The classifier improved its classification accuracy by adding weight to distance in the classification phase. Because single classifier uses single feature and has its limitation, improved distance metric can't get expected result.It's suggested that different classifiers offered complementary information. So how to combine classifiers is the key of improving recognition performance. We do some researches on multiple classifiers combination. Multiple classifiers combination method is composed of cascade connection and parallel connection, a deep research about two methods is presented. At first, classifiers ensemble method based on adaptive distance metric via Bagging technique is proposed.Furthermore, through analyzing minimum distance classifier and support vector machine, a new method of combining classifiers is proposed, which make full use of advantages of minimum distance classifier and support vector machine. Firstly, the algorithm computes the distance from the sample to the pattern classification, and finds the two smallest distances, then defines a confidence. If the confidence is greater than the classification threshold, the MDC can give the final result; otherwise, the SVM would be training by samples, which are belong to the two smallest distances, the sample would be classified on SVM.In the thesis, experiment is tested on UCI standard database, the experimental results show that the proposed methods are effective.