Study Of Multiple Classifiers Combination

This paper studies the design of pattern recognition system based multiple classifiers combination. At the aspect of classifier ensemble design, two methods are proposed. The first means define a correlation function based on the Weighted Misrecognition Probability (WMRP) as the criterion to select classifiers. The second means uses genetic algorithms to select the best pattern property subset to compose classifier ensemble. Experiment results show both of the two methods have a higher accuracy of recognition. Especially when the pattern to classify has some properties that are bad for classification, the latter has very good performance. An improved neural network combination method in the pattern recognition system of multiple classifier combination is also proposed to increase the classification correctness. The experiment result shows the means has a better performance than the traditional neural network combination method and the classification means by a classifier. Finally, a method based on clustering is proposed to realize the classifier selection procedure in the pattern recognition system based on multiple classifier selection. Experiment result proves its availability, and shows its better performance on pattern recognition than some combination methods and all of the individual classifiers when each of the used classifiers is an expert for a species of pattern.